Machine Learning-Based Prediction of Dividend-Increase Behavior from Corporate Financial Indicators and Its Application to Portfolio Management

Abstract

This study examines whether corporate information supplements management dividend forecasts in predicting dividend increases and whether the predictions improve stock selection in Japan. Using financial and market indicators, management forecasts, and dividend histories available at each prediction date, we develop machine learning models for Tokyo Stock Exchange firms with March fiscal year-ends, excluding banks. A rolling training, validation, and retraining procedure produces out-of-sample predictions for 2013–2023, with portfolio evaluation from June 2013 to May 2024. We compare forecast-based benchmarks and multivariate models. Then, we combine predicted probabilities with forecast dividend yields. Multivariate eXtreme Gradient Boosting (XGBoost) achieves a mean annual area under the receiver operating characteristic curve of 0.86, compared with 0.77 for XGBoost using only the forecast dividend yield change. The combined long–short strategy achieves an annualized Sharpe ratio of 1.26 and a positive alpha under the Fama–French three-factor model (t = 4.12, p < 0.001). Its arithmetic cumulative spread return remains positive under the examined transaction cost scenarios. These findings suggest that dividend increase predictions can complement forecast dividend yields in stock selection, whereas practical implementation requires consideration of borrowing costs and other trading constraints not separately assessed here.

1. Introduction

In recent years, listed firms have been increasingly called upon to improve their capital efficiency. Figure A1 in Appendix A separately shows the percentages of firms with a return on equity (ROE) below 8% and with a price-to-book ratio (PBR) below 1, highlighting challenges in enhancing firm value and returning capital to shareholders. Against this background, in March 2023, the Tokyo Stock Exchange (TSE) called on all companies listed on the Prime and Standard Markets to implement management practices conscious of the cost of capital and stock prices.

Dividend increases and share repurchases have attracted attention as measures for improving capital efficiency. Dividend increases, in particular, are shareholder payout measures that firms are more likely to choose when they are confident in stable future earnings and cash flows; they are therefore regarded as important decisions reflecting a firm’s stance and outlook. In recent years, moves to strengthen shareholder payouts through dividend increases have accelerated amid public concern about the accumulation of retained earnings and market calls for improved capital efficiency.

Dividend increases also serve multiple functions beyond returning funds to shareholders, including the appropriate use of surplus funds, adjustment of capital structure, and signaling a firm’s future outlook to the market. From a theoretical perspective, Jensen (1986) noted that excessive free cash flow within a firm may lead managers to undertake inefficient reinvestments, thereby eroding shareholder value. Returning funds to shareholders through dividend increases is an effective way to mitigate such agency problems. In addition, the asymmetric-information models of Bhattacharya (1979) and Miller and Rock (1985) show that dividend policy functions as a signaling mechanism that indirectly conveys information held within the firm to the market.

Dividend increases have also attracted attention because of their effects on stock prices. Julio and Ikenberry (2004) show that the market responds favorably to resumed or increased dividends. In Japan, Uchiyama and Takahashi (2015) document pronounced positive stock-price reactions to dividend-increase announcements.

These reactions suggest that dividend increases function as a clear positive signal in the Japanese market, where firms tend to manage dividend policy with a long-term emphasis on stability.

In contrast, share repurchases are a major shareholder payout measure but have a decision-making structure different from that of dividend increases. In particular, although dividend increases involve rigidity in that “a dividend level, once raised, cannot easily be lowered,” share repurchases tend to be implemented flexibly in response to business cycles and stock-price levels (Jagannathan et al., 2000). Consequently, dividend increases reflect long-term judgments about firms’ permanent earning capacity, whereas share repurchases are often based on short-term factors, such as stock-price undervaluation or temporary surplus funds. Another distinctive feature of dividend increases is that firms disclose forecast dividends in advance. Forecast dividends provide important information to investors, but managers tend to set them conservatively, and actual dividends frequently exceed forecast amounts. Therefore, forecast dividends alone may not fully explain the occurrence of dividend increases.

Mochizuki et al. (2026) show that share repurchase behavior can be predicted using firm-level financial, market, governance, and historical repurchase information and that the resulting predicted probabilities can be applied to stock selection. They provide a framework in which a shareholder payout decision is first predicted and the prediction is then evaluated as a stock selection signal. However, in the case of dividend increases, two unresolved issues remain. First, firms disclose forecast dividends before the realized dividend is known; therefore, investors already observe a management-based signal of future dividend decisions. Consequently, it is unclear whether firm-level information provides additional predictive information beyond management dividend forecasts. Second, forecast dividend yield is already a conventional stock selection measure based on expected dividend income. Even if future dividend increases can be predicted, it remains unclear whether the predicted probability provides information that complements this existing measure. These two unresolved issues motivate the following two research questions of this study.

First, to what extent can machine learning models combining financial and market indicators with historical dividend information improve the prediction of future dividend increases relative to approaches based on management dividend forecasts? Second, can combining predicted dividend increase probabilities with forecast dividend yields improve portfolio performance relative to strategies using either measure alone?

To address the first research question, we augment the financial, market, and governance indicators used in the share repurchase prediction framework with dividend-specific information, including historical dividend changes, positive and negative dividend surprises, and historical dividend yields. We compare the resulting multivariate models with management-forecast-based benchmarks and use feature importance and SHapley Additive exPlanations (SHAP) values to examine the information contributing to the predictions. To address the second research question, we combine predicted dividend increase probabilities with forecast dividend yields and compare the resulting strategy with strategies using either measure alone. Thus, this study extends the prior prediction framework by examining the incremental predictive information specific to dividend increases and its potential value in dividend-yield-based stock selection. The analysis concerns predictive associations and does not identify causal effects underlying dividend increase decisions.

2. Literature Review

Building on prior research on the determinants of dividend-increase behavior and its effects on stock prices, this study seeks to predict dividend-increase behavior using machine learning and applies the resulting predictions to equity portfolio management. This section reviews the relevant literature in four categories: theoretical foundations common to dividends and share repurchases, distinctive characteristics of and research on share repurchases, distinctive characteristics of and research on dividend increases, and related research using machine learning.

2.1. Common Theoretical Foundations of Dividends and Share Repurchases

Dividends and share repurchases are shareholder payout measures that share two principal theoretical foundations: agency theory and signaling theory. First, Jensen’s (1986) free cash flow hypothesis plays a central role in agency theory. This hypothesis argues that when excessive funds accumulate within a firm, managers may direct them toward inefficient investment or organizational expansion, potentially impairing shareholder value. Therefore, dividends and share repurchases are mechanisms for distributing surplus funds outside the firm and reducing agency costs. Recent evidence is consistent with this agency-based interpretation. Kim et al. (2024) find that dividend payments reduce managerial bad news hoarding and overinvestment, providing empirical evidence that dividend payouts can constrain managerial behavior associated with agency problems.

Second, according to Bhattacharya (1979) and Miller and Rock (1985), signaling theory emphasizes that dividends and share repurchases can convey a firm’s private information about future earning capacity. Strengthening shareholder payouts can therefore communicate managerial confidence and earnings outlooks to investors. Chang et al. (2024) examine how management forecasting ability affects the information conveyed by dividends. They find that the positive association between dividend increases and subsequent earnings improvements among Japanese firms is stronger when managers have greater earnings forecasting ability. This finding indicates that the information about future earnings contained in dividend changes depends on management forecasting ability.

Moreover, these theoretical arguments have empirical support. Grullon et al. (2002) and Uchiyama and Takahashi (2015) report positive abnormal returns following dividend increase announcements. Ikenberry et al. (1995) report medium- to long-term excess returns following share repurchase announcements. Furthermore, recent studies provide evidence that both forms of shareholder payout are associated with positive stock market reactions. Michałowski (2025) reports a statistically significant positive abnormal return of 1.39% for dividend increases on the Warsaw Stock Exchange. Similarly, Anolick et al. (2021) report positive cumulative abnormal returns for a large sample of open market share repurchases across nine European countries. Despite these commonalities, dividends and share repurchases differ substantially in their decision-making structures and implementation timing. The next subsection summarizes the characteristics specific to share repurchases and clarifies how they differ from dividend policy.

2.2. Distinctive Characteristics of Share Repurchases and Related Research

Like dividends, share repurchases are shareholder payout measures, but their decision-making processes have distinctive characteristics. Jagannathan et al. (2000) point out that share repurchases can be adjusted flexibly in response to business cycles and stock-price levels, whereas dividends exhibit rigidity because firms are expected to maintain a dividend level once it has been increased. Share repurchases therefore tend to be used for more flexible purposes, such as responding to temporary cash surpluses or correcting perceived stock-price undervaluation. Arena and Julio (2023) provide evidence of this flexibility: firms facing greater litigation risk reduce dividends and shift payout toward share repurchases, whereas cash-constrained firms reduce repurchases when settlement costs are realized.

Empirical studies have confirmed that share repurchases are sensitive to external conditions and market valuations. Vermaelen (1981) shows that share repurchases function as a signal of firm undervaluation and reports that repurchase activity increases when firms perceive their shares to be undervalued. Focusing on changes in ownership structure, Kitamura and Kojima (2019) show that, for regional banks, the probability of a share repurchase rises as foreign institutional ownership increases. This finding provides an example of shareholder monitoring pressure and governance demands influencing managerial behavior and supports the view that share repurchases are affected not only by market valuation but also by ownership structure. More recently, Bourveau et al. (2023) show that institutional-investor behavior can also affect repurchase decisions through market valuation: selling by short-term-oriented mutual funds can create undervaluation, to which portfolio firms respond by increasing share repurchases.

Mochizuki et al. (2026) use machine-learning models to predict share-repurchase behavior among Japanese firms and calculate feature importance to analyze factors underlying the decision-making process. According to the feature importance results, “share repurchase yield (past year)” is the most important variable, followed by variables capturing prior share-repurchase activity. These results suggest strong persistence in share-repurchase behavior. “Foreign ownership ratio,” “market capitalization,” “equity ratio,” and “PBR” also show non-negligible importance, indicating that external-investor pressure, market valuation, and financial soundness are relevant to the decision. Furthermore, they apply the predicted share repurchase probabilities directly to stock selection by constructing portfolios ranked by the model outputs. This provides a prediction-to-portfolio framework that this study extends using the predicted probability as an additional screening dimension (rather than as the sole stock selection signal) within a strategy based on forecast dividend yield.

Together, the prior literature suggests that share-repurchase decisions are often based on short-term factors that depend on the external environment. By contrast, dividend increases appear to reflect more long-term, structural judgments; the next subsection summarizes their distinctive characteristics.

2.3. Distinctive Characteristics of Dividend Increases and Related Research

Although dividend increases and share repurchases are both shareholder payout measures, their decision-making structures differ fundamentally. First, dividend policy exhibits strong rigidity because firms are expected to maintain a dividend level once it has been increased. Lintner (1956) shows that firms tend to base dividend-increase decisions on permanent earning capacity rather than temporary profits, indicating that such decisions are made cautiously and from a long-term perspective. Lin and Yu (2025) examine managerial characteristics associated with downward dividend stickiness. They find that overconfident CEOs increase dividends more when earnings rise but reduce them less when earnings fall, linking asymmetric dividend adjustments to managerial overconfidence. Drawing on life-cycle theory, DeAngelo et al. (2006) report that mature firms tend to favor dividend increases as retained earnings accumulate and growth opportunities decline. Additionally, Akindayomi and Amin (2022) provide recent evidence that firms’ longer-term strategic characteristics are related to dividend increase decisions. They find that innovation-oriented firms are less likely to increase dividends than cost-efficiency-oriented firms, with the difference being more pronounced among growth firms. Taken together, these studies suggest that dividend increases are associated with persistent earnings capacity, managerial characteristics, corporate life cycle conditions, and longer-term strategic orientation, consistent with the view that dividend increase decisions are relatively rigid and structural compared with more flexible payout measures such as share repurchases.

Second, a notable feature of dividend increases in the Japanese market is the existence of forecast dividends. Forecast dividends are the dividend amounts that a firm’s management expects to pay, based on its assessment of the firm’s condition and prospects. However, managers tend to set forecast dividends conservatively in light of future uncertainty, and actual dividends commonly exceed forecast amounts when financial results are finalized. It therefore remains unclear to what extent forecast dividends reflect managers’ private information, and forecast values alone have limited ability to explain the occurrence of dividend increases. Understanding dividend-increase decisions thus requires consideration of diverse factors not fully captured by forecast dividends, including complex relationships among financial indicators, corporate governance structures, and firms’ stages of growth. However, much of the existing literature focuses on stock-price reactions following dividend-increase announcements or descriptions of firm characteristics, leaving scope to examine whether firm-level information provides incremental predictive value for dividend increases beyond management dividend forecasts. Moreover, prior studies rely mainly on linear models and have not fully captured nonlinear interactions among financial indicators or the complexity of corporate governance structures. Accordingly, understanding dividend-increase behavior requires an integrated view of multiple information sources, including institutional features such as dividend rigidity and forecast dividends, as well as financial indicators, firm attributes, and market valuation.

2.4. Related Research Using Machine Learning

In recent years, a growing body of research has used firm characteristics and market information to predict stock returns and applied these predictions to portfolio management. Gu et al. (2020) showed that machine-learning estimates of risk premiums can provide greater predictive power and economic value for investment than conventional regression models. Their study demonstrates that extracting information embedded in firm characteristics through machine learning can bridge asset-pricing research and practical stock selection.

However, direct prediction of stock returns presents a distinct challenge. Lo (2004) proposes the adaptive markets hypothesis, in which market efficiency changes as market participants learn and compete, suggesting that the stock-price formation mechanism is not fixed. Gu et al. (2020) also showed that simply using a large number of explanatory variables can reduce out-of-sample predictive performance through overfitting, necessitating regularization or dimensionality reduction. These findings indicate that low signal-to-noise ratios and model generalizability are central challenges in direct return predictions.

In response to these challenges, recent studies have incorporated corporate-event information into machine-learning models for stock-market prediction and portfolio construction. Wu (2020) uses event sequences contained in financial news to predict movements in stock indices and individual stocks and shows that event information is useful for understanding market behavior and improving investment performance. Ye and Schuller (2020) also analyze post-earnings-announcement drift using XGBoost and show that machine-learning models built around corporate events can be applied to portfolio construction. These studies support the use of corporate-event information in investment models, although their prediction targets are stock-price movements rather than the occurrence of corporate events.

Michaely et al. (1995) showed that dividend initiations and omissions have substantial effects on stock prices.

Taken together, prior research shows that corporate payout actions can be treated as prediction targets and that the resulting predictions can be applied to stock selection. However, in the case of dividend increases, management dividend forecasts already provide an observable benchmark for prediction, whereas forecast dividend yield is already used as a conventional dividend-based stock selection measure. Therefore, the remaining gap pertains to whether dividend-specific firm information provides incremental predictive value beyond management forecasts and whether the resulting predicted probability can add stock selection value when used as an additional screening dimension within a strategy based on forecast dividend yield. These two issues define the contribution of this study.

Table 1 summarizes recent studies relevant to the shareholder payout mechanisms and prediction framework reviewed above.

Table 1. Recent research relevant to shareholder payout mechanisms and prediction (2021–2026).

Theme and Study	Research Focus and Approach	Main Finding	Relevance to this Study
Agency costs—Kim et al. (2024)	Empirical analysis of dividend payments and managerial behavior associated with agency problems.	Dividend payments reduce managerial bad news hoarding and overinvestment.	Provides recent empirical support for the agency-based view that dividend payouts can constrain managerial behavior.
Signaling—Chang et al. (2024)	Japanese dividend-paying firms; analysis of dividend changes, subsequent earnings, and management forecasting ability.	The positive association between dividend increases and subsequent earnings improvements is stronger when managers have greater forecasting ability.	Explains a condition affecting the information about future earnings conveyed by dividend changes.
Market evidence—Michałowski (2025)	Study of dividend increases and cuts on the Warsaw Stock Exchange from 2015 to 2024.	A statistically significant positive average abnormal return of 1.39% is associated with dividend increases.	Provides recent evidence of positive stock market reactions associated with dividend increases.
Market evidence—Anolick et al. (2021)	Analysis of 882 open market share repurchases in nine European countries.	A positive cumulative average abnormal return of 1.37% is associated with open market share repurchases.	Provides recent evidence of positive stock market reactions associated with share repurchases.
Payout flexibility—Arena and Julio (2023)	Analysis of litigation risk and payout choices among U.S. firms.	Firms facing greater litigation risk reduce dividends and shift payout toward repurchases; cash-constrained firms reduce repurchases when settlement costs are realized.	Provides recent evidence of repurchase flexibility in response to litigation-related cash needs.
Ownership and valuation—Bourveau et al. (2023)	Analysis of institutional investor trading, market valuation, and firms’ repurchase responses.	Short-term-oriented mutual fund selling can create undervaluation, to which portfolio firms respond by increasing repurchases.	Shows a channel through which institutional investor behavior and market valuation affect repurchase decisions.
Repurchase prediction—Mochizuki et al. (2026)	Machine learning prediction of share repurchases among Japanese firms, followed by portfolio evaluation.	Historical repurchase information is prominent in the prediction models, and predicted probabilities are applied to stock selection.	Provides the baseline prediction-to-portfolio framework. This study extends it by introducing dividend-specific information, evaluating incremental predictive value relative to management dividend forecasts, and using predicted dividend increase probability as an additional screening dimension within a strategy based on forecast dividend yield.
Dividend stickiness—Lin and Yu (2025)	Analysis of managerial overconfidence and dividend adjustments to earnings changes.	Overconfident CEOs increase dividends more when earnings rise but reduce them less when earnings fall.	Relates asymmetric dividend adjustments to managerial characteristics.
Business strategy—Akindayomi and Amin (2022)	Examines business strategy and firms’ decisions to pay, initiate, and increase dividends.	Innovation-oriented firms are less likely to increase dividends than cost-efficiency-oriented firms, and the difference is stronger among growth firms.	Provides recent evidence that dividend increase decisions are related to persistent firm-level strategic and growth characteristics.

Note. This table summarizes recent research supporting the theoretical background and research positioning of Section 2.1, Section 2.2 and Section 2.3.

Figure 1 summarizes how the theoretical background informs the empirical analysis. Agency and signaling considerations, together with dividend rigidity and corporate life cycle characteristics, motivate the use of firm-level financial, market, ownership, and historical dividend information to predict future dividend increases.

Figure 1. Conceptual framework of dividend increase prediction and its application to stock selection. Note. The framework links the theoretical background, information available at the prediction date, and the two research questions. A supplementary analysis using compounded cumulative returns compares the three portfolio strategies under annual rebalancing. The arrows indicate information use and evaluation links.

The empirical analysis is organized around the two research questions stated in Section 1. Section 4 evaluates the predictive information these inputs provide beyond management dividend forecasts through out-of-sample classification and model interpretation. Section 5 uses portfolio performance, risk factor adjustment, and transaction cost scenarios to evaluate whether combining predicted dividend increase probabilities with forecast dividend yields improves stock selection relative to using either measure alone.

3. Trend Analysis of Shareholder Payout Measures

This section describes the shareholder payout behavior underlying the prediction and portfolio analyses. After defining the sample and data, we document the annual incidence of dividend increases and share repurchases, examine annual stock price return patterns around dividend increase years, and compare management dividend forecasts with realized dividends. These analyses provide the empirical context for assessing dividend increase predictability in Section 4 and its potential use in portfolio management in Section 5.

3.1. Sample and Data Overview

The sample consisted of firms listed on the TSE First Section and, following the April 2022 market segment restructuring, firms listed on the TSE Prime Market. We further restricted the sample to firms with fiscal years ending in March and recorded information available at the end of May each year. Firms in the TOPIX-17 “Banks” category were excluded from the sample, whereas non-bank financial firms were not excluded as a category. The sample period extended from the end of May 2011 to the end of May 2024, and an annual panel dataset was constructed using financial information, market information, and company forecasts observed at the end of May each year.

The sample and information reference dates were selected for three reasons. First, to continuously observe firms with relatively high liquidity and a high likelihood of belonging to institutional investors’ investment universes throughout the sample period, the analysis covered TSE First Section listed firms before the April 2022 market segment restructuring and TSE Prime Market listed firms thereafter. Under the TSE’s listing criteria, these market segments require sufficient market capitalization and liquidity for firms to be potential investment targets for many institutional investors, as well as high governance standards. Restricting the sample to these market segments was therefore expected to provide a reasonable degree of investability and cross-firm comparability. Second, banks were excluded because their balance sheet structure, leverage, and financial intermediation activities differ substantially from those of ordinary nonfinancial firms, reducing the comparability of several firm-level financial indicators used in the prediction model. Mohanram et al. (2018) show that some conventional financial statement indicators are unavailable or not directly applicable to banks and develop bank-specific indicators for stock selection. More recently, Katsafados et al. (2026) use bank-specific accounting information in machine learning models for predicting bank stock returns. The industry controls in this study follow the TOPIX-17 classification, which treats “Banks” as a separate category from “Financials (ex Banks).” Therefore, nonbank financial firms were retained in the sample and treated through the corresponding industry classification. Finally, most listed firms close their fiscal years in March, and their full-year results are released primarily between late April and mid-May each year. The sample was restricted to firms with March fiscal year-ends, and the end of May each year was designated as the information reference date to place firms on a common information date and reduce differences arising from financial results that have not yet been disclosed. This approach allowed stocks to be selected using a common information set after the latest full-year results and after company forecasts had been disclosed for most firms, enabling portfolio management to begin on the first business day of June.

Market data, including stock prices, and firm-level financial and major shareholder information used in this study are available through Refinitiv Eikon. Management earnings and dividend forecasts are disclosed in corporate earnings releases and subsequent revisions through TDnet. The analysis uses local copies of the relevant data. Stock prices and the occurrence of dividend increases were recorded at each month-end, whereas financial statement information was recorded using a common reference date of the end of May each year.

The 2011–2023 annual datasets contain 13,172 base sample observations and 13,165 observations in the final machine learning sample. Annual sample counts and dividend increase rates are reported in Table A1, and the missing-value treatment is described in Appendix B.2.

3.2. Incidence of Shareholder Payout Measures

Dividend increases and share repurchases are part of firms’ capital policies, but their incidence and patterns may be influenced by firm characteristics and the macroeconomic environment. Characterizing these patterns before constructing a prediction model is useful for understanding the phenomena under analysis. This subsection presents the annual evolution of shareholder payout measures and summarizes their overall patterns.

Figure 2 shows the annual incidence rates of dividend increases. The incidence rate is calculated by dividing the number of firms that increased dividends in each fiscal year by the total number of sample firms; for example, the FY2011 value is based on dividend increases implemented between the first business day of June 2011 and the end of May 2012. From 2011 to 2023, the proportion of firms that increased dividends rose sharply, from 33.6% to 72.1%. This rise may partly reflect progress in corporate reform following the TSE’s March 2023 call for improved capital efficiency and revisions to listing standards. By contrast, the incidence of dividend increases fell from 53.3% in FY2018 to 41.9% in FY2019 and recovered only partially to 45.8% in FY2020. This decline may be attributable to firms retaining surplus funds to prepare for unforeseen contingencies during the outbreak of a novel infectious disease.

Figure 2. Annual incidence rate of dividend increase behavior [%] (number of dividend-increasing firms/total number of sample firms).

Figure 3 presents the annual incidence rates for combinations of shareholder payout measures. Between 2011 and 2023, the proportion of firms undertaking neither measure fell sharply from 55.1% to 19.2%. Against the backdrop of the TSE’s call for improved capital efficiency and revisions to listing standards, this pattern suggests that firms have become more proactive toward shareholder payouts. In FY2023, 72.1% of firms increased dividends, whereas only 37.5% conducted share repurchases. These figures indicate that dividend increases are more commonly selected than share repurchases as a shareholder payout measure in the Japanese market. Moreover, the proportion of firms implementing both dividend increases and share repurchases rose from 8.0% in 2011 to 28.8% in 2023, indicating a growing tendency to combine the two rather than rely on a single payout measure. This pattern is consistent with firms strategically using shareholder payout measures to improve capital efficiency.

Figure 3. Annual incidence rate of shareholder payout measures [%] (number of firms in each category/total number of sample firms).

3.3. Annual Stock Price Return Patterns Around Dividend Increase Years

This subsection examines the relationship between realized dividend increases and annual stock price returns, focusing on whether mean market-adjusted returns in the dividend increase year are positive and comparing return patterns for all dividend-increasing firms with those for firms that had not increased dividends in the preceding year. Section 2 provides the background for examining the relationship between dividend increases and stock performance. This subsection organizes the data by dividend increase year rather than announcement date and provides a retrospective comparison of annual returns.

Figure 4 presents annual stock-price change rates for all selected dividend-increasing firms from three years before to three years after the dividend-increase year, which is defined as relative year 0.

Figure 4. Stock price change rates before and after dividend increases (market-average adjusted). The horizontal axis shows relative year i, with the dividend-increase year t defined as the event year (i = 0). The vertical axis shows the mean adjusted stock-price change rate (%) defined in Equation (2). The market average is the sample-based equal-weighted benchmark described in Section 3.3. Each thin line represents an event year t; the thick line and shaded band show, respectively, the mean and ±1 sample standard deviation across available event-year observations at each relative year.

For each event year t, the market benchmark is constructed from firms with a TOPIX indicator of 1 in the corresponding annual dataset. The TOPIX field is used as a membership indicator, not as an index-return series. The benchmark universe is fixed at event year t, and its equally weighted mean return is calculated for each relative year using available observations. The adjusted return is the mean return of the selected dividend-increasing firms minus this benchmark. For Figure 5, the additional restriction is applied only to the dividend-increasing firms; the benchmark remains unchanged.

Figure 5. Same as Figure 4, but restricted to firms that had not increased dividends during the preceding year.

Let $U_t$ denote the benchmark universe selected in event year t and let $C_t$ denote the dividend-increasing firms selected for Figure 4 or Figure 5. For each relative year i ∈ {−3, −2, −1, 0, 1, 2, 3}, let $U_{t,i}$ and $C_{t,i}$ denote the firms in these sets with an available annual return in year t + i, respectively. Let $N_{t,i}$ be the number of firms in $C_{t,i}$, and let $R_{k,t+i}$ denote the annual stock-price return of firm k in that year. The equal-weighted benchmark return is defined as follows:

$$R_{\mathrm{A},t,i}={\displaystyle \frac{1}{\left|U_{t,i}\right|}}\sum _{k\in U_{t,i}}{R}_{k,t+i}$$

The firm-level adjusted return and its mean across the selected dividend-increasing firms are then defined by Equations (1) and (2):

$${R\prime }_{k,t,i}=R_{k,t+i}-R_{\mathrm{A},t,i}$$

(1)

$$\mathrm{E}\left[{R\prime }_{t,i}\right]={\displaystyle \frac{1}{N_{t,i}}}{\sum }_{k\in C_{t,i}}{R\prime }_{k,t,i}$$

(2)

The mean notation $\mathrm{E}\left[{R\prime }_{t,i}\right]$ denotes the sample mean of the adjusted returns for event year t and relative year i. In Figure 4 and Figure 5, each thin line represents one event year from 2013 to 2023. The thick line is the mean across available event-year observations at each relative year, and the shaded band is the mean plus or minus one sample standard deviation. Missing observations are excluded rather than treated as zeros.

According to Table 2, the mean adjusted annual stock price return across dividend increase years in Figure 4 is 3.08% at relative year 0 and 5.21% at relative year −1. Thus, the selected firms exhibit positive average annual returns relative to the sample-based benchmark in both the dividend-increase year and the preceding year.

Table 2. Results of One-Sided t-Tests of Whether the Mean Adjusted Stock-Price Change Rate Is Significantly Positive in Figure 4 (Unrestricted Sample) and Figure 5 (Restricted Sample).

Relative Year	n	Mean [%] (Unrestricted)	Standard Deviation [%] (Unrestricted)	t-Statistic (Unrestricted)	Mean [%] (Restricted)	Standard Deviation [%] (Restricted)	t-Statistic (Restricted)
−3	11	1.18	2.3	1.7	−3.28	2.88	−3.77
−2	11	2.29	1.61	4.71 **	−6.84	2.76	−8.22
−1	11	5.21	1.87	9.27 **	−0.4	4.1	−0.32
0	11	3.08	2.06	4.97 **	4.48	2.19	6.80 **
1	10	0.08	1.88	0.14	−0.46	3.92	−0.37
2	9	−0.31	1.04	−0.9	−0.94	2.63	−1.07
3	8	−0.19	1.75	−0.31	2.72	3.08	2.50 *

Note. Relative year 0 denotes the dividend increase year, −1 the preceding year, and +1 the following year. The unrestricted sample includes all selected dividend-increasing firms; the restricted sample includes only those that had not increased dividends in the preceding year. For each relative year, the observations used in the one-sided one-sample t-test are the annual group mean market-adjusted returns for the available dividend increase years from 2013 to 2023 (Equation (2)). The reported mean and standard deviation summarize these annual group means. Thus, n counts dividend-increase-year observations (not individual firms). For each relative year, n is the same in the two samples; it is 11 for relative years −3 through 0 and declines to 10, 9, and 8 for relative years +1, +2, and +3 because later returns extend beyond the dataset. The null hypothesis is that the population mean is nonpositive, and the alternative is that it is positive. * and ** denote one-sided p-values below 0.05 and 0.01, respectively.

To examine whether annual return patterns vary with recent dividend history, Figure 5 restricts the dividend-increasing sample to firms that had not increased dividends in the preceding year while retaining the same market benchmark. In the dividend increase year (i = 0), the mean adjusted return is 4.48% in the restricted sample and 3.08% in the unrestricted sample, a difference of 1.40 percentage points. In the preceding year (i = −1), the corresponding means are −0.40% and 5.21%, respectively, indicating that the positive return pattern in the preceding year becomes substantially less pronounced under this restriction. The restriction concerns only whether a firm increased dividends in the preceding year and does not identify the first dividend increase in its history. The differences between the sample means are descriptive comparisons; the t-tests in this subsection compare each mean with zero rather than directly testing the difference between the two samples.

For each relative year i, we conduct a one-sided one-sample t-test using the annual mean adjusted returns for event years 2013–2023 with available observations. The null hypothesis is that the population mean is non-positive, and the alternative is that it is positive.

Here, i denotes the time position relative to the dividend increase year. Each observation in the test sample is the group mean adjusted return for one dividend increase year t. Accordingly, holding i fixed, the test is conducted across event-year observations rather than by pooling the individual returns of all firms.

$$H_0:{\mu }_i\le 0, H_1:{\mu }_i>0$$

$$T_i={\displaystyle \frac{{\overline{x}}_i}{\frac{s_i}{\sqrt{n_i}}}}$$

(3)

Here, $n_i$ is the number of non-missing event-year observations, ${\overline{x}}_i$ is their sample mean, and $s_i$ is their sample standard deviation calculated with $n_i-1$ degrees of freedom. The one-sided p-value is obtained from the upper tail of a t distribution with $n_i-1$ degrees of freedom.

Table 2 summarizes the test results. Under the one-sided tests specified above, the mean adjusted annual stock price return in the dividend increase year (i = 0) is significantly positive at the 1% level in the unrestricted and restricted samples. The results show a positive association between realized dividend increases and stock price performance relative to the sample-based benchmark during the same year within the sample and period examined. As firms are grouped using realized dividend outcomes, this subsection does not identify announcement date effects or establish the profitability of an ex ante tradable strategy. The usefulness of dividend increase predictions for stock selection is evaluated separately in Section 5 using out-of-sample predictions.

For example, in the restricted sample, −1 denotes the year preceding a dividend increase. The corresponding t-statistic of −0.32 is associated with a mean adjusted return of −0.40%; therefore, the null hypothesis of a nonpositive mean is not rejected. At relative year 0, the mean adjusted return is 4.48%, and the t-statistic of 6.80 ** indicates a significantly positive mean at the 1% level. These tests compare each mean with zero; they do not test the difference between the restricted and unrestricted samples.

3.4. Trend Analysis of Management Forecasts

This subsection examines the gap between management dividend forecasts and realized dividends, focusing on how often forecasts are subsequently revised and whether forecasted dividend changes correspond to realized dividend outcomes. This comparison provides empirical context for the first research question: whether firm-level information can supplement management dividend forecasts in predicting dividend increases.

First, Figure 6 shows how forecast dividends disclosed by firms in each fiscal year were subsequently revised. An upward revision is defined as a case in which the actual dividend exceeds the forecast dividend, a downward revision as a case in which the actual dividend falls below the forecast, and no change as a case in which the forecast and actual dividends are equal. Although the proportion of observations classified as upward revisions fluctuated between FY2011 and FY2023, it showed an upward long-term trend, reaching approximately 75% in FY2023. This result indicates that an increasing share of firms set forecast dividends conservatively and subsequently raise dividends after financial results are finalized. Thus, although forecast dividends provide important information for investors, they do not fully explain the extent to which firms ultimately increase shareholder payouts.

Figure 6. Annual incidence rates of revisions to forecast dividends [%] (number of firms in each revision category/number of firms with classifiable forecast revisions).

Next, to examine the consistency between management forecasts and actual dividends more directly, we cross-tabulate all observations from FY2011 to FY2023. The results are summarized in Table 3. The matrix shows how dividend policies initially announced by management, namely an increase, no change, or a cut, correspond to actual dividend outcomes. Notably, cases in which firms forecast no change or a dividend cut but ultimately increased dividends accounted for approximately 24% of all observations. There were 2759 cases in which a forecast of no change was followed by an actual dividend increase and 399 cases in which a forecast dividend cut was followed by an actual increase. Together, these observations were classified as “unexpected dividend increases.” This finding is consistent with the tendency of Japanese firms to set forecast dividends conservatively and shows that forecast dividends alone do not fully capture dividend-increase behavior.

Table 3. Cross-Tabulation of Management Forecasts and Actual Dividends (FY2011–FY2023).

Actual/Forecast	Increase	No Change	Decrease
Increase	4043 (31%)	2759 (21%)	399 (3%)
No Change	128 (1%)	4322 (33%)	225 (2%)
Decrease	112 (1%)	443 (3%)	741 (6%)

Overall, firms’ dividend forecasts contain useful information about future dividend policy, but non-negligible discrepancies from actual dividends remain. Predictions of dividend increases based solely on management forecasts are therefore limited, and financial information, market information, ownership information, and historical dividend records may contain incremental predictive information beyond forecast dividends. Accordingly, this study examines the extent to which dividend increases not captured by forecast dividends can be predicted by incorporating these sources of information into a multivariate nonlinear framework.

4. Predictability of Dividend-Increase Behavior Using Machine Learning

This section describes the machine-learning models constructed to predict the occurrence of dividend increases. Specifically, it describes the data design, model selection and optimization, evaluation metrics, and visualization methods used to interpret the results, and then examines the extent to which machine learning can predict dividend-increase behavior.

4.1. Machine Learning

The prediction target (target variable) in this study is whether a firm increases its dividends during the following year, and the task is formulated as a binary classification problem. As in the analyses in the preceding sections, the prediction sample is restricted to firms listed on the TSE First Section or TSE Prime Market with fiscal years ending in March. The final modeling sample is obtained using the missing value treatment described in Appendix B.

For the explanatory variables, we first adopted the financial, market, and governance indicators used by Mochizuki et al. (2026) in their share-repurchase prediction models (Table 4). In addition, focusing on the characteristics of dividend-increase behavior, we introduced new variables related to firms’ historical dividend-policy outcomes. Specifically, as shown in Table 5, we added “historical dividend record,” “dividend-surprise history,” and “historical dividend yields.” The share-repurchase indicator (next year), shown in the first row of Table 4 as the target variable in the prior framework, is neither an explanatory variable nor the target variable in the present model.

Table 4. Explanatory variables based on Mochizuki et al. (2026). Their original share-repurchase target (first row) is shown for reference only. Standardized scores are calculated cross-sectionally; industry indicators follow Section 3.1.

Variable	Type
share-repurchase indicator (next year)	0 (No) or 1 (Yes)
Δ forecast dividend yield [%]	Continuous value (0 indicates no change in dividend)
forecast dividend yield [%]	Standardized score
log(market capitalization)	Standardized score
forecast cash-flow-to-price ratio	Standardized score
cash-rich ratio	Standardized score
forecast after-tax return on total assets [%]	Standardized score
forecast return on equity (ROE) [%]	Standardized score
forecast operating margin [%]	Standardized score
forecast return on assets (ROA) growth rate [%]	Standardized score
forecast ROE growth rate [%]	Standardized score
forecast five-year ordinary profit growth rate [%]	Standardized score
current ratio [%]	Standardized score
fixed assets-to-long-term capital ratio [%]	Standardized score
equity ratio [%]	Standardized score
interest-bearing debt-to-total-assets ratio [%]	Standardized score
debt-to-total capital ratio [%]	Standardized score
interest-bearing debt-to-sales ratio [%]	Standardized score
60-day return (3-month return)	Standardized score
240-day return (1 year)	Standardized score
240-day volatility (1 year)	Standardized score
240-day variance (1 year)	Standardized score
one-year share turnover ratio	Standardized score
foreign ownership ratio	Standardized score
individual investor ownership ratio	Standardized score
TOPIX-17 industry indicator (excluding Banks)	0 (Not applicable) or 1 (Applicable)
ROE	Standardized score
PBR	Standardized score
dividend payout ratio	Standardized score
share-repurchase indicator (past year)	0 (No) or 1 (Yes)
share-repurchase indicator (1–2 years prior)	0 (No) or 1 (Yes)
share-repurchase indicator (2–3 years prior)	0 (No) or 1 (Yes)
share repurchase yield (past year)	Continuous value
share repurchase yield (1–2 years prior)	Continuous value
share repurchase yield (2–3 years prior)	Continuous value

Table 5. Target variable (first row) and newly introduced explanatory variables (all other rows) for the dividend-increase prediction model. These explanatory variables are combined with those in Table 4, excluding its first-row share-repurchase target.

Variable	Type
dividend-increase indicator (next year)	0 (No) or 1 (Yes)
dividend-increase indicator (past year)	0 (No) or 1 (Yes)
dividend-increase indicator (1–2 years prior)	0 (No) or 1 (Yes)
dividend-increase indicator (2–3 years prior)	0 (No) or 1 (Yes)
dividend-cut indicator (past year)	0 (No) or 1 (Yes)
dividend-cut indicator (1–2 years prior)	0 (No) or 1 (Yes)
dividend-cut indicator (2–3 years prior)	0 (No) or 1 (Yes)
positive dividend surprise indicator (past year)	0 (No) or 1 (Yes)
positive dividend surprise indicator (1–2 years prior)	0 (No) or 1 (Yes)
positive dividend surprise indicator (2–3 years prior)	0 (No) or 1 (Yes)
negative dividend surprise indicator (past year)	0 (No) or 1 (Yes)
negative dividend surprise indicator (1–2 years prior)	0 (No) or 1 (Yes)
negative dividend surprise indicator (2–3 years prior)	0 (No) or 1 (Yes)
dividend yield (past year)	Continuous value
dividend yield (1–2 years prior)	Continuous value
dividend yield (2–3 years prior)	Continuous value

The “historical dividend record” comprises variables indicating year-on-year changes in actual dividends, classified as an increase, no change, or a decrease, over the preceding 1 to 3 years. As Lintner (1956) notes, dividend policy exhibits strong rigidity, and firms tend to increase dividends only when they are confident that earnings have improved on a permanent basis. Recent dividend records may therefore provide important information for predicting future dividend-increase behavior.

“Dividend-surprise history” indicates whether actual dividends exceeded or fell below the firm’s prior dividend forecasts. Japanese firms tend to set dividend forecasts conservatively, and positive dividend surprises, in which actual dividends exceed forecast amounts, occur in a non-negligible number of cases. Such positive dividend surprises may reflect management confidence in future performance that was not explicitly conveyed in the disclosed forecasts or a less conservative disclosure stance. The forecast–actual dividend discrepancy may also capture managerial caution, forecast accuracy, and disclosure practices. Accordingly, dividend-surprise history supplements aspects of managerial decision-making not fully captured by dividend forecasts and may help identify latent information relevant to dividend-increase decisions.

These additional variables were adopted because the empirical findings in the preceding section indicate that dividend forecasts alone cannot fully explain the occurrence of dividend increases. Their inclusion was intended to improve dividend-increase prediction by incorporating historical dividend-policy information and managerial disclosure behavior.

The machine-learning models adopted for the binary classification problem were k-nearest neighbors, a support vector machine, logistic regression, a random forest, and XGBoost (eXtreme Gradient Boosting).

Each model was trained and evaluated sequentially using a sliding window procedure designed to prevent leakage of future information. The training and evaluation design follows Mochizuki et al. (2026), and Figure 7 provides the authors’ own illustration of how this procedure is implemented in this study. Firm-level information available at the end of May each year provided the explanatory variables, and whether each firm increased its dividend by the end of the following May was used as the prediction target. We refer to the data pairing the input information as of the end of May in year X with the dividend increase outcome from the first trading day of June in year X through the end of May in year X + 1 as the “year-X dataset.” The procedure is illustrated below using 2011–2013 datasets as an example.

Figure 7. Sliding-window machine-learning and performance-evaluation procedure with the training data rolled forward annually. Source: Authors’ own illustration. The training and evaluation design follows Mochizuki et al. (2026).

• Training step: Optuna was used to specify a candidate set of hyperparameters, and the model was trained using the 2011 dataset. This procedure was repeated for multiple candidate hyperparameter sets.
• Validation step: The model trained in Step 1 was applied to the 2012 dataset, and its generalization performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC). The hyperparameter combination with the highest AUC was selected from the candidates.
• Performance and portfolio-evaluation step: The best hyperparameters identified in Step 2 were fixed, and the model was retrained on the 2012 dataset. The 2013 dataset was then used to obtain out-of-sample predictions. Final model performance was evaluated and portfolio simulations were conducted using these out-of-sample predictions.

By repeating Steps 1–3 while rolling the window forward by 1 year at a time, we generated out-of-sample predictions for the 2013–2023 datasets, covering whether firms increased dividends from the first trading day of June 2013 through the end of May 2024.

The one-year training window follows the rolling design used in the related prediction framework and emphasizes recent cross-sectional relationships. Using several preceding years would increase the number of training observations; however, it would also combine observations from more distant periods. As the relationship between firm characteristics and payout decisions may change over time, older observations may contain patterns that are less relevant to the subsequent prediction period. Therefore, we use the most recent annual cross-section for model fitting. This design prioritizes recency but does not imply that one year is the optimal training window length.

Appendix C provides the software and implementation details.

Generative AI was used as an auxiliary tool for Japanese-to-English translation, manuscript revision, and code review. AI-assisted suggestions were checked and revised by the authors before being incorporated into the manuscript or analysis code.

4.2. Evaluation of Classification Performance

Table 6 reports the classification performance of each machine-learning model on the out-of-sample data for the 2013–2023 datasets. Performance was evaluated separately for each year, and the table reports averages over the full period. The evaluation metrics were AUC, accuracy, precision, recall, and the F1 score. For probability-based machine-learning classifiers, accuracy was calculated by applying a threshold of 0.5 to the predicted probability of a dividend increase.

Table 6. Classification performance of the dividend-increase prediction models (annual averages).

Machine-Learning Model	AUC	Accuracy	Recall	Precision	F1 Score
Δ forecast dividend yield (sign-based classification)	―	0.73	0.57	0.95	0.71
XGBoost (Δ forecast dividend yield only)	0.77	0.72	0.61	0.88	0.72
k-Nearest Neighbors	0.77	0.69	0.78	0.71	0.74
Support Vector Machine	0.82	0.74	0.75	0.80	0.77
Logistic Regression	0.83	0.76	0.76	0.82	0.79
Random Forest	0.86	0.76	0.74	0.83	0.78
XGBoost	0.86	0.77	0.76	0.83	0.79

Bold values indicate the numerical maximum for each metric, not statistical significance.

Precision is the proportion of firms predicted to increase dividends that actually do so, whereas recall is the proportion of actual dividend-increasing firms correctly identified. The precision values in Table 6 are the averages of the annual test sample proportions; for example, 0.95 corresponds to a mean annual precision of 95% rather than a single proportion calculated after pooling all years. These values summarize out-of-sample predictive performance. Numerical differences between the reported averages are descriptive comparisons and are not presented as statistically significant differences between models.

The “forecast dividend yield” explanatory variable in Table 4 is the latest publicly available dividend forecast observable at the end of May each year divided by the share price at that same date. “Δ forecast dividend yield” is the forecast dividend yield minus the most recent actual dividend yield; a positive value indicates that management forecasts a dividend increase, zero indicates no change, and a negative value indicates a dividend decrease. The actual dividend yield is calculated using the same share price as the forecast dividend yield to ensure a common basis.

Because directly referring to Δ forecast dividend yield is the most basic approach available to an investor seeking to predict dividend increases, we use it as a benchmark for the machine-learning models in this study. Specifically, a positive Δ forecast dividend yield is classified as a dividend increase, whereas a value of zero or less is classified as no dividend increase. In addition to this sign-based classification, an XGBoost model using only Δ forecast dividend yield as an explanatory variable is included as another benchmark.

Table 6 presents the classification performance of these approaches. For the sign-based classification using Δ forecast dividend yield, precision was very high at 0.95, whereas recall was 0.57. Thus, this benchmark produced few false-positive dividend-increase predictions but missed many firms that subsequently increased dividends. One possible explanation is the asymmetry in the consequences of forecast errors: investors may react negatively if management forecasts a dividend increase that is not realized, whereas an unanticipated dividend increase may be received favorably. This asymmetry may encourage conservative management forecasts. Consequently, Δ forecast dividend yield derived from these forecasts may also be conservative and miss many firms that ultimately increase dividends. In the XGBoost model using only Δ forecast dividend yield, recall improved to 0.61, but limitations due to insufficient information remained. By contrast, all multivariate machine-learning models exhibited predictive ability across model types. XGBoost performed particularly well, with an AUC of 0.86, precision of 0.83, recall of 0.76, and an F1 score tied for the highest among the models. These results suggest that jointly using financial indicators, market information, ownership information, historical dividend records, and dividend-surprise information provides incremental predictive information beyond Δ forecast dividend yield alone. Appendix C reports the Optuna hyperparameter search ranges, number of trials, and best hyperparameters selected for each year.

Figure 8 and Figure 9 show the classification characteristics of XGBoost, one of the best-performing models listed in Table 6. The left panel of Figure 8 shows the ROC curves for the XGBoost model using only Δ forecast dividend yield, whereas the right panel shows those for the multivariate XGBoost model using all explanatory variables. The ROC curve visualizes overall classification performance independently of a particular classification threshold through the relationship between the true-positive and false-positive rates. In the left panel, the ROC curves bow toward the upper left corner to some extent even when only Δ forecast dividend yield is used, confirming that this indicator contains useful predictive information about future dividend-increase behavior. In the right panel, the ROC curves for the multivariate XGBoost model lie more consistently toward the upper-left corner, and the AUC is approximately 0.86. This result indicates that simultaneously using multiple sources of information, including financial indicators and past dividend behavior, in addition to the direct information on dividend changes contained in Δ forecast dividend yield, further improves discrimination between firms with and without a dividend increase. Figure 9 shows the frequency distribution of XGBoost outputs by class label, indicating that firms with higher model outputs were more likely to increase dividends.

Figure 8. Annual receiver operating characteristic curves and area-under-the-curve values obtained from XGBoost. The (left) panel uses only Δ forecast dividend yield as an explanatory variable, whereas the (right) panel uses all explanatory variables.

Figure 9. Frequency distribution of XGBoost outputs by class label (whether the firm increased dividends during the following year).

Overall, the multivariate machine-learning models, particularly XGBoost, extracted diverse information related to dividend-increase behavior and showed useful predictive performance.

4.3. Interpretation and Visualization of Explanatory Variables

As XGBoost was one of the best-performing models in the preceding comparison, we used its fitted results to visualize the importance of the explanatory variables and their contributions to the target variable.

Figure 10 and Figure 11 report the feature importance of each explanatory variable, measured by the extent to which it was used in tree splits. As Figure 10 shows, Δ forecast dividend yield is the most important variable, followed by variables related to forecast–actual discrepancies and profitability, including the positive dividend surprise indicator (past year), forecast after-tax return on total assets, and forecast return on equity (ROE). Although the dividend-increase indicator (past year) also has some importance, the positive dividend surprise indicator (past year) has greater feature importance. This suggests that, beyond a prior dividend increase itself, the “surprise” component, that is, an actual dividend above the firm’s prior forecast, provides particularly strong predictive information for future dividend-increase behavior. Because firms tend to announce conservative dividend forecasts, actual dividends above those forecasts may indicate strong management confidence in corporate performance and a proactive stance toward shareholder payouts. The model therefore appears to place greater weight on forecast–actual discrepancies, particularly positive dividend surprises, than on the simple historical record of dividend increases.

Figure 10. Explanatory variables with the highest feature importance in XGBoost.

Figure 11. Explanatory variables with the lowest feature importance in XGBoost.

In addition, stock-price return and payout-related indicators, such as the 240-day return (1 year), 60-day return (3-month return), and dividend payout ratio, show relatively high importance. This pattern is consistent with prior findings that firms’ financial conditions and stock-price movements are related to dividend-increase decisions and suggests that the machine-learning model extracts economically plausible features from the data.

By contrast, as Figure 11 shows, industry indicators have very low importance. This suggests that dividend-increase behavior depends more strongly on dynamic factors, such as forecast–actual discrepancies, corporate profitability, and market valuation, than on fixed attributes such as industry.

Next, SHAP values were calculated to interpret the direction and magnitude of each explanatory variable’s contribution to the model output. SHAP values show the extent to which each explanatory variable raises the model output (positive contribution) or lowers it (negative contribution) for individual predictions. In this study, SHAP values were calculated by applying the model trained in each year to the out-of-sample data for the following year, and values obtained over the full period were pooled to produce Figure 12 and Figure 13. The explanatory variables on the vertical axis are arranged from top to bottom in descending order of mean absolute SHAP values.

Figure 12. Distribution of SHapley Additive exPlanations values in the XGBoost model, for the top-ranked explanatory variables with large contributions to the model output.

Figure 13. Same as Figure 12, but for the lower-ranked explanatory variables with small contributions to the model output.

Among the top-ranked variables in Figure 12, higher values (red points) of Δ forecast dividend yield, 240-day return (1 year), and positive dividend surprise indicator (past year) are associated with larger positive SHAP values, indicating positive contributions to the model output. By contrast, lower values (blue points) of forecast dividend yield and actual price-to-book ratio (PBR) tend to be associated with larger positive SHAP values. The recent institutional context is relevant to interpreting this pattern. In March 2023, the TSE called on all companies listed on the Prime and Standard Markets to implement management practices conscious of the cost of capital and stock prices. In response, many firms have faced pressure to adopt management practices that are more conscious of the cost of capital, and undervalued firms with low PBRs may have stronger incentives to increase shareholder payouts through dividends or share repurchases. The SHAP finding that low PBR shifts the predicted probability toward a dividend increase is consistent with this institutional context. However, the pooled SHAP results include years preceding this request and should not be interpreted as identifying its effect.

Among the lower-ranked variables in Figure 13, industry indicators such as “Transportation and Logistics” and variables on past share-repurchase activity, such as the share-repurchase indicator (past year) and the share-repurchase indicator (1–2 years prior), have SHAP values concentrated approximately between −0.5 and 0.5. Therefore, these variables appeared to have little influence on the prediction model.

Each point represents the SHAP value for one firm. The horizontal axis shows the contribution to the model output, and the color indicates the magnitude of the explanatory variable.

To examine variable-level relationships in greater detail, Figure A2 and Figure A3 in Appendix A compare observed dividend increase rates with mean XGBoost predictions across groups of explanatory variable values. The grouping procedure is described alongside the figures. The model predictions generally track the observed incidence patterns, including nonlinear relationships for some variables. The principal observations from these visualizations are as follows.

• Model Output

The model outputs (predicted probabilities) closely matched the observed values (incidence rates), indicating that XGBoost captured the observed patterns.

• Dividend-Related Variables

The dividend-increase indicator (past year) shows a clear tendency for firms with a recent dividend increase to increase dividends again in the following year. This pattern indicates strong persistence in dividend policy as a form of corporate behavior and is consistent with the “dividend rigidity” identified by Lintner (1956).

Δ forecast dividend yield also shows a strong positive association with dividend increases. Because Japanese firms generally set dividend forecasts conservatively, firms that nevertheless announce a forecast dividend increase may be more likely to implement one, reflecting management confidence in future earning capacity.

The “dividend-surprise history” newly introduced in this study provides useful information for explaining dividend-increase behavior. In particular, firms with a positive dividend surprise in the past year have a markedly higher probability of increasing dividends in the following year, indicating that firms whose actual dividends exceeded their prior forecasts are more likely to maintain a proactive stance toward shareholder payouts. By contrast, negative dividend surprises over the preceding 1 to 3 years are associated with lower dividend-increase probabilities, and firms that previously paid dividends below forecasts tend to adopt more cautious dividend policies. Thus, dividend-surprise information provides information about both the conservatism and accuracy of management forecasts.

For both dividend yield (past year) and forecast dividend yield, firms with higher yields have lower probabilities of increasing dividends in the following year. This may reflect greater managerial caution because firms already offering high dividend yields have limited capacity for further dividend increases.

• Share-Repurchase-Related Variables

No clear relationship is observed between share repurchase yield and the probability of a dividend increase, although firms that conducted a share repurchase in the past year show a slight tendency to increase dividends more frequently. This pattern may reflect a stronger commitment to shareholder payouts and greater financial capacity. Accordingly, although past share repurchases are not a direct determinant of dividend increases, they may function as a complementary signal of a firm’s stance toward shareholder payouts.

• Firm Size

A clear monotonic positive relationship is observed between market capitalization and the probability of dividend increases. As Ueno and Baba (2005) note, this pattern may reflect the characteristics of mature firms, which have more stable cash flows and fewer investment opportunities and therefore may find it easier to distribute excess funds to shareholders as dividends.

• Profitability and Growth

Indicators of firms’ earning capacity, including forecast after-tax return on total assets, forecast return on equity (ROE), and forecast operating margin, show a positive relationship with the probability of a dividend increase. This pattern is consistent with the financial constraint that dividend increases must be supported by firms’ earning capacity and suggests that more profitable firms are more likely to increase dividends.

For growth indicators, such as the forecast return on assets (ROA) growth rate, forecast ROE growth rate, and forecast five-year ordinary profit growth rate, the probability of a dividend increase is highest near the mean standardized score of zero and declines as the score moves away from zero, revealing an inverted-U-shaped nonlinear pattern. This pattern is consistent with corporate life-cycle theory (DeAngelo et al., 2006). Specifically, Myers and Majluf (1984) note that firms in the growth stage with exceptionally high growth expectations tend to allocate internal funds preferentially to business expansion, thereby constraining shareholder payouts. At the other extreme, as Easterbrook (1984) suggests, mature firms with markedly low growth rates have a weakening earnings base from which to fund dividend increases and may find it difficult to secure sufficient dividend capacity. These considerations imply that dividend increases are most likely among mature firms that maintain profitability and have fewer large-scale investment opportunities. This result suggests that dividend-increase behavior is related to a firm’s stage in the corporate life cycle.

• Financial Soundness and Financial Capacity

For the current ratio, the probability of a dividend increase tends to be highest at an intermediate level, perhaps because an excessively low ratio indicates a shortage of funds, whereas an excessively high ratio suggests inefficient cash management. This pattern suggests that firms may be more likely to make shareholder payouts when they maintain appropriate liquidity while using funds effectively.

Distinctive patterns were also observed for indicators related to financial structure. Firms with higher fixed assets-to-long-term capital ratios are less likely to increase dividends, indicating that as more funds are tied up in fixed assets, funding flexibility declines and the capacity to allocate funds to shareholder payouts becomes more limited. The probability of a dividend increase also declines as the interest-bearing debt-to-total-assets ratio increases, suggesting that dividend capacity may be constrained when debt repayment is prioritized. By contrast, firms with higher equity ratios are more likely to increase dividends, indicating that financial stability is an important factor supporting shareholder payouts.

These results are consistent with Easterbrook’s (1984) argument that a firm’s financial capacity plays an important role in payout behavior. In other words, firms require sufficient financial capacity and a stable capital structure to increase dividends, and firms with sound financial foundations are more likely to pursue an active dividend policy.

• Market Valuation

For PBR, firms with standardized scores below zero tend to have a higher probability of increasing dividends. This interpretation is consistent with the emphasis on capital efficiency in the TSE’s 2023 request to all companies listed on the Prime and Standard Markets (Tokyo Stock Exchange, 2023). Undervalued firms may face greater capital-market pressure and may respond by strengthening shareholder payout measures, including dividend increases, to improve market valuation.

Firms with higher 60-day returns (3-month returns) and 240-day returns (1 year) are more likely to increase dividends. Strong stock-price performance may coincide with favorable earnings outlooks and, in turn, a more proactive stance toward shareholder payouts. Accordingly, the higher probability of dividend increases among high-return firms may reflect a more optimistic future outlook.

For indicators of market uncertainty and trading activity, including 240-day volatility (1 year), 240-day variance (1 year), and the one-year share turnover ratio, dividend increases tend to be most frequent when standardized scores are near zero. At extremely high levels, large stock-price fluctuations, excessive trading activity, or unstable market conditions may impede a stable dividend policy. By contrast, at extremely low levels, limited market attention may make firms less likely to commit to shareholder payouts. Thus, the greater frequency of dividend increases under relatively stable market conditions may indicate that firms emphasize stability when determining dividend policy.

• Ownership Structure

The tendency for firms with higher foreign ownership ratios to have higher probabilities of dividend increases is consistent with Kitamura and Kojima (2019) and suggests that shareholder pressure emphasizing capital efficiency may play a role. By contrast, the probability of a dividend increase tends to decline as the individual investor ownership ratio rises, suggesting that management may prioritize long-term business investment over short-term measures to support stock prices. The debt-to-total capital ratio is also negatively associated with the probability of a dividend increase, suggesting that firms with lower debt proportions in their capital structures, and correspondingly higher equity proportions, tend to favor dividend increases as shareholder payout measures.

5. Applicability to Equity Portfolio Management

5.1. Portfolio Construction and Performance

Section 4 demonstrated the predictability of dividend-increase behavior. The retrospective association documented in Section 3 motivates an examination of whether dividend increase predictions are useful for stock selection. This subsection constructs equity portfolios using out-of-sample predicted probabilities of dividend increases and evaluates their performance.

This section calculates the predicted probability of a dividend increase for each firm (stock) and constructs portfolios based on that probability. Specifically, as shown in Figure 7 of the preceding section, the hyperparameters are optimized, and the model parameters are retrained using a sliding-window procedure before the model is applied to new out-of-sample data. Repeating this process annually produces out-of-sample dividend-increase probabilities at the end of May in each year from 2013 to 2023, with the corresponding one-year evaluation periods ending in May 2014 through May 2024.

As in the preceding section, the explanatory variables consist of publicly available information at the end of May each year. This timing was used because information on firms with fiscal years ending in March generally becomes widely available in the market around May of the same year. To account for this information lag, the machine-learning model is updated at the end of May each year, and the equity portfolios are rebalanced based on the resulting predicted probabilities of dividend increases. Each new portfolio begins trading on the first business day of June and is held for 1 year through the end of the following May. Based on the predicted probability for each stock, all stocks are sorted into terciles, and an equally weighted long portfolio is constructed for each tercile. For comparison, we also construct a long portfolio from stocks that increased dividends during the prediction period. However, because these stocks are unknown on the prediction date, this portfolio represents an idealized state that cannot be traded in practice. We refer to it as the “perfect-foresight” portfolio. In addition, returns are calculated for a long–short portfolio formed by shorting stocks in the bottom tercile of model outputs and taking long positions in stocks in the top tercile, thereby examining the usefulness of predicted probabilities for stock selection.

As the analysis in this section corresponds to active management, performance is evaluated using excess returns relative to the Tokyo Stock Price Index (TOPIX), which serves as the passive-management benchmark. Specifically, let $R_{p,m}$ denote the return on portfolio p in month m, and $R_{\mathrm{M},m}$ denote the total return on TOPIX. The excess return on a long portfolio, ${R\prime }_{p,m}$, is then defined as follows:

$${R\prime }_{p,m}=R_{p,m}-R_{\mathrm{M},m}$$

(4)

The monthly values are then arithmetically summed, as in ${\sum }_{m=1}{R\prime }_{p,m}$, and the resulting series is termed the “cumulative excess return”; its time-series evolution is plotted. For a long–short portfolio, let p₁ denote the long portfolio in the top tercile and p₂ denote the long portfolio in the bottom tercile.

$${R\prime }_{p,m}=(R_{p1,m}-R_{\mathrm{M},m})-(R_{p2,m}-R_{\mathrm{M},m})=R_{p1,m}-R_{p2,m}$$

(5)

Because the TOPIX return cancels in Equation (5), it need not be subtracted when calculating ${R\prime }_{p,m}$. Thus, for a long–short portfolio, ${R\prime }_{p,m}=R_{p,m}$; Equation (5) is referred to in this paper as the “long–short spread return.”

Figure 14 plots the cumulative excess returns and long–short spread returns of portfolios formed on the predicted probability of a dividend increase. Portfolio performance is broadly rank-ordered as follows: perfect foresight > top predicted-probability tercile > middle predicted-probability tercile > bottom predicted-probability tercile, indicating the usefulness of the predicted probability. In particular, returns for stocks with low model outputs (the bottom predicted-probability tercile) remain consistently weak throughout the period, whereas stocks with high predicted probabilities generate stable excess returns relative to the market. This result indicates that the predicted probability of a dividend increase provides a practically useful signal for stock selection.

Figure 14. Cumulative excess returns and long–short spread returns of portfolios formed on predicted dividend-increase probabilities.

By contrast, a widely used traditional investment approach based on dividend policy is investing in high-dividend-yield stocks. However, a high dividend yield does not necessarily indicate a firm’s future growth or dividend sustainability, and such strategies may tilt toward mature or low-growth firms (DeAngelo et al., 2006). Consequently, stock selection based solely on dividend yield risks selecting firms whose yields are temporarily high but whose underlying earning capacity and scope for further dividend increases are limited.

To address these limitations of the traditional approach, this study proposes using the output of the dividend-increase prediction model developed in Section 4, namely the predicted probability of a future dividend increase. Whereas forecast dividend yield measures the dividend income implied by management forecasts relative to the current share price, the predicted probability of a dividend increase provides a forward-looking signal of dividend growth over the following year, based on financial indicators, growth measures, dividend history, and dividend-surprise information. In this sense, the two measures capture the level of dividend income and the prospect of dividend growth, respectively, and may therefore provide complementary information for stock selection.

To examine this complementarity, Figure 15 groups firms by forecast-dividend-yield terciles and realized dividend outcomes—increase, no change, or decrease—and compares average monthly returns (%) across the resulting nine cells without subtracting TOPIX. Average returns differ substantially by forecast dividend yield and realized dividend outcome. Firms with high forecast dividend yields and realized dividend increases have the highest returns, whereas low-yield firms that cut dividends have the lowest returns, supporting the importance of jointly evaluating these two dimensions. These findings suggest that jointly considering the “level” and “change” of dividends may identify groups of stocks with greater investment efficiency.

Figure 15. Differences in average monthly returns (%) by forecast dividend yield and realized dividend outcome.

The combined strategy first divides firms into forecast-dividend-yield terciles, then selects the highest-probability third within the highest-yield tercile and the lowest-probability third within the lowest-yield tercile for the long and short legs, respectively. Each leg therefore contains approximately one ninth of the eligible stocks. The following three investment strategies are compared to evaluate the usefulness of this approach.

(a) Traditional high-dividend-yield strategy: takes long positions in stocks in the top 33% of forecast dividend yields and short positions in stocks in the bottom 33%.

(b) Dividend-increase-prediction-only strategy: takes long positions in stocks in the top 33% of predicted probabilities and short positions in stocks in the bottom 33%. This strategy evaluates the standalone usefulness of the machine-learning dividend-increase prediction model and is equivalent to the results shown in Figure 14.

(c) Combined dividend-increase-prediction and forecast-dividend-yield strategy: considers both the predicted probability and forecast dividend yield, taking long positions in the cell in which both the predicted probability of a future dividend increase and forecast dividend yield are high and short positions in the cell in which both are low. This strategy aims to improve stock-selection efficiency by combining dividend level with future dividend growth.

Figure 16a–c plot the cumulative excess returns and long–short spread returns of the three strategies. First, the traditional high-dividend-yield strategy (Figure 16a) shows a recent recovery in its spread return; over the long run, however, its performance is unstable, and sustained returns are difficult to obtain. The annualized Sharpe ratio of its spread return is 0.71. The dividend-increase-prediction-only strategy (Figure 16b) has generated a positive spread return over the long run, but its recent performance has weakened. The annualized Sharpe ratio of its spread return is 0.44. By contrast, the combined strategy (Figure 16c) has generated a stable long-term spread return and has the highest annualized Sharpe ratio, at 1.26. The spread-return comparison in Figure 16d also shows that the combined strategy consistently outperforms the other strategies, indicating that information integration improves portfolio efficiency.

Figure 16. Comparison of portfolio management based on the dividend-increase prediction model and forecast dividend yield. (a) Cumulative excess returns and long–short spread returns for an investment strategy representing a conventional high-dividend-yield fund; (b) cumulative excess returns and long–short spread returns for the strategy based solely on dividend-increase predictions; (c) cumulative excess returns and long–short spread returns for the strategy combining dividend-increase predictions and forecast dividend yields; (d) long–short spread returns for the three strategies shown in panels (a–c).

Appendix D provides an additional robustness check based on compounded cumulative returns. Using the same portfolio selections and monthly return inputs as the main analysis, each portfolio is equally weighted at the annual June formation and then held without within-year rebalancing, allowing constituent weights to evolve with realized returns. The year-end portfolio value is carried into the following annual formation. Under this specification, the probability-sorted portfolios retain their broad performance ordering, and the combined strategy exhibits the largest terminal compounded cumulative return differential (CCRD) between the top and bottom long-only portfolios among the three strategies. These results provide supplementary support for combining forecast dividend yield with the predicted probability of a dividend increase.

5.2. Portfolio Performance After Risk-Factor Adjustment

The preceding subsection confirmed the usefulness of the combined strategy, which uses the predicted probability of a dividend increase together with forecast dividend yield. However, if the returns obtained can be explained by established risk factors, such as market risk, the small-cap effect, and the value effect, the incremental contribution of the strategy would be limited. Accordingly, this subsection regresses the monthly returns of each strategy on the Fama–French three-factor model and examines whether positive returns remain after risk-factor adjustment. We use the Fama–French three-factor model as a conventional benchmark for an initial risk-adjusted evaluation of portfolio performance. We examine whether the portfolio returns remain positive after controlling for market, size, and value exposures. This specification is not intended to exhaustively explain the sources of portfolio returns; therefore, the resulting alpha should be interpreted as conditional on the three-factor benchmark. Evaluation using more recent and richer multifactor models, together with stricter portfolio evaluation settings, remains an important robustness extension for future research. Specifically, we estimate the following regression for the dependent return for each portfolio p:

$$Y_{p,m}={\alpha }_p+{\beta }_p^{\mathrm{M}\mathrm{K}\mathrm{T}}{\mathrm{M}\mathrm{K}\mathrm{T}}_m+{\beta }_p^{\mathrm{S}\mathrm{M}\mathrm{B}}{\mathrm{S}\mathrm{M}\mathrm{B}}_m+{\beta }_p^{\mathrm{H}\mathrm{M}\mathrm{L}}{\mathrm{H}\mathrm{M}\mathrm{L}}_m+{\epsilon }_{p,m},$$

(6)

where $Y_{p,m}$ is the dependent return for portfolio p in month m, ${\mathrm{M}\mathrm{K}\mathrm{T}}_m$ is a market factor, ${\mathrm{S}\mathrm{M}\mathrm{B}}_m$ is a size factor, and ${\mathrm{H}\mathrm{M}\mathrm{L}}_m$ is a value factor. The factor data were obtained from French (2025). ${\mathrm{M}\mathrm{K}\mathrm{T}}_m$ is the market portfolio return net of the risk-free rate. In addition, α represents the additional return not explained by the three risk factors, and ${\epsilon }_{p,m}$ is the error term.

For the top and bottom stock portfolios, the dependent variable is the portfolio return net of the risk-free rate. For the long–short spread (SP), the dependent variable is the top-minus-bottom return, without a further risk-free-rate deduction.

Coefficients are estimated by ordinary least squares. Statistical inference uses Newey–West heteroskedasticity- and autocorrelation-consistent standard errors with a maximum lag of three months, and p-values are two-sided.

Table 7a–c present the results. Focusing first on the long–short portfolio’s spread return (SP), the βHML coefficient is large and statistically significant for the traditional high-dividend-yield strategy, indicating strong exposure to the value factor. One possible reason is that lower stock prices produce higher dividend yields and, as shown in the SHAP analysis in Figure 12, lower PBRs are associated with higher probabilities of dividend increases; accordingly, undervalued firms may be more inclined to pursue dividend payout policies. By contrast, the HML coefficient is 0.125 for the combined strategy, compared with 0.406 for the traditional high-dividend-yield strategy, indicating a smaller estimated positive exposure to the value factor. The combined strategy’s HML loading nevertheless remains significant at the 5% level.

Table 7. Tests of Alpha Significance Using the Fama–French Three-Factor Model.

(a) Significance tests for strategies based solely on dividend-increase predictions
	Top Predicted-Probability Tercile		Bottom Predicted-Probability Tercile		SP (Top − Bottom)
	Coefficient	t-value	Coefficient	t-value	Coefficient	t-value
α	0.265	1.14	0.009	0.04	0.257	2.16 *
βMKT	0.813	8.80 ***	0.792	8.69 ***	0.021	0.56
βSMB	0.213	1.63	0.406	3.13 **	−0.193	−3.68 ***
βHML	0.165	1.93	0.435	4.89 ***	−0.270	−5.49 ***
(b) Significance Tests for the Investment Strategy Representing a Conventional High-Dividend-Yield Strategy
	Top Forecast-Dividend-Yield Tercile		Bottom Forecast-Dividend-Yield Tercile		SP (Top − Bottom)
	Coefficient	t-value	Coefficient	t-value	Coefficient	t-value
α	0.263	1.27	0.018	0.08	0.245	3.14 **
βMKT	0.809	9.27 ***	0.809	8.70 ***	0.000	0.00
βSMB	0.376	2.97 **	0.269	2.01 *	0.107	2.13 *
βHML	0.487	5.71 ***	0.081	0.96	0.406	13.59 ***
(c) Significance Tests for the Combined Strategy
	Top Forecast-Dividend-Yield Tercile and Top Predicted-Probability Tercile		Bottom Forecast-Dividend-Yield Tercile and Bottom Predicted-Probability Tercile		SP (Top − Bottom)
	Coefficient	t-value	Coefficient	t-value	Coefficient	t-value
α	0.407	1.73	−0.263	−1.15	0.670	4.12 ***
βMKT	0.833	8.95 ***	0.838	7.99 ***	−0.005	−0.10
βSMB	0.308	2.25 *	0.426	2.90 **	−0.117	−1.28
βHML	0.443	4.89 ***	0.319	3.50 ***	0.125	2.11 *

Note: *, **, and *** indicate p < 0.05, p < 0.01, and p < 0.001, respectively. SP is the top-minus-bottom return and is not subject to a further risk-free-rate deduction.

We next focus on α for the risk-factor-adjusted long–short spread return (SP). All three strategies have positive alphas that are statistically significant, at the 5% level for the dividend-increase-prediction-only strategy, the 1% level for the traditional high-dividend-yield strategy, and the 0.1% level for the combined strategy. The combined strategy has the largest estimated alpha, 0.670 (t = 4.12), suggesting that its spread return includes a positive component not fully explained by market, size, and value factors.

5.3. Portfolio Turnover and Transaction Costs

To further evaluate the combined strategy from a practical perspective, this subsection examines how portfolio turnover from rebalancing and transaction costs affect portfolio performance. In this study, transaction costs are assumed to be proportional to the value traded, and each stock’s position weight is defined as its position value relative to a fixed reference notional. At each annual June formation, stocks are equally weighted within each leg. During the subsequent one-year holding period, changes in individual stock prices alter position values and their weights even without trading. Accordingly, annual turnover is estimated from the difference between the new annual target weights and the drifted position-value weights immediately before rebalancing. The annual portfolio turnover (%) at the end-of-May rebalance is calculated using the following equation:

$$\begin{gathered} \mathsf{\Delta }{w}_y=100{\sum }_s{\sum }_i\left|w_{i,y}^s-{\tilde{w}}_{i,y}^s\right| \\ {w}_{i,y}^s=\left\{\begin{array}{ll}{\displaystyle \frac{1}{N_{\mathrm{L},y}}}& s=\mathrm{L},i\in L_y\\ -{\displaystyle \frac{1}{N_{\mathrm{S},y}}}& s=\mathrm{S},i\in S_y\\ 0& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}\right. \end{gathered}$$

(7)

More specifically, the drifted position-value weight immediately before the next annual rebalance is calculated by multiplying the initial target weight by the product of the 12 monthly gross-return factors (one plus each monthly stock return) over the holding period:

$${\tilde{w}}_{i,y}^s=w_{i,y-1}^s{\displaystyle {\prod }_{m=1}^{12}}\left(1+r_{i,y-1,m}\right)$$

Here, i indexes the stocks, and y denotes the portfolio formation year. For Figure 17a, the long and short legs are each equally weighted and scaled to unit absolute weights, with $N_{\mathrm{L},y}$ and $N_{\mathrm{S},y}$ denoting the numbers of stocks in the respective legs. The stock weights are defined as above. For each leg, the target position-value weight is zero for stocks not selected into that leg. In Equation (7), s denotes the long (L) or short (S) leg; $w_{i,y}^s$ is the target signed position-value weight for the new formation year, and ${\tilde{w}}_{i,y}^s$ is the signed position-value weight immediately before rebalancing. In the drift formula, $r_{i,y-1,m}$ denotes the decimal monthly return of stock i in month m of the preceding holding year. Position values are updated each month without within-year rebalancing, and their weights remain measured relative to the fixed reference notional rather than being renormalized to the year-end portfolio value. The absolute trading amounts of the long and short legs are calculated separately and then summed. Therefore, turnover reflects the difference between the new target and the drifted pre-rebalancing position values, rather than changes in successive annual target weights. Initial formation is measured relative to zero stock positions, giving turnover of 200%. After the initial formation, even complete replacement of all constituent stocks need not give turnover of exactly 400%, because the values of positions held over the preceding year may have changed.

Figure 17. Practical analysis of the combined strategy (long–short portfolio). Panel (a) uses unit absolute target weights for each of the long and short legs relative to the fixed reference notional; panel (b) uses the unscaled long-minus-short spread with transaction costs based on the same turnover measure.

Figure 17a reports turnover on a 100% long/100% short absolute-notional basis. The long- and short-leg contributions are shown separately, and their sum gives the overall turnover. The initial value of 200% represents portfolio establishment from zero positions. In subsequent years, turnover varies with changes in constituents and position values during the holding period, with higher values at the 2020 and 2021 rebalances.

Next, transaction costs are calculated on the basis of portfolio turnover and deducted from the combined strategy’s long–short spread return $R_{p,m}$. Let c denote the transaction cost rate in percentage units, with $c\in \{0.00,\hspace{0.25em}0.10,\hspace{0.25em}0.30,\hspace{0.25em}0.50,\hspace{0.25em}0.80\}$.

We examine transaction cost rates of 0.10%, 0.30%, 0.50%, and 0.80%, together with a zero-cost benchmark. Zurek and Heinrich (2021) assume a cost of 0.10% per unit of traded value in a study including Japanese equities, whereas Yamamoto et al. (2020) adopt a one-way cost of 0.50% for Japanese equity portfolios. Together, these two assumptions motivate the range considered here. The 0.80% rate is included as a conservative high-cost stress scenario rather than as an estimate of a market-wide average.

The assumed rates represent simplified transaction-cost scenarios proportional to traded value; stock-borrowing and other holding-period-dependent expenses are not estimated separately.

In particular, this analysis does not establish whether each short position can be initiated and maintained under the borrowing conditions applicable to that stock. Moreover, applying a common cost rate per unit of traded value does not provide separate estimates of security-specific execution costs. The 0.80% scenario is a stress test of the assumed trading cost component, not an upper bound on all implementation expenses. Therefore, the positive cumulative spread in Figure 17b should be interpreted conditional on these assumptions rather than as an investor-level return net of all implementation expenses.

As rebalancing occurs once per year, the transaction cost is divided by 12, allocated across months, and deducted from $R_{p,m}$. Annual costs are allocated evenly across the 12 holding months for this sensitivity analysis; this allocation does not represent the actual timing of cost payments.

Let $T_y^{\mathrm{L}}$ and $T_y^{\mathrm{S}}$ denote the turnover of the long and short legs, respectively, each normalized to unit absolute weight. The spread return is calculated as $R_{\mathrm{L},m}$ − $R_{\mathrm{S},m}$ without an additional 0.5 multiplier. The sum of the two leg turnovers is the overall turnover shown in Figure 17a and is used directly to estimate transaction costs.

$$\Delta w_y=T_y^{\mathrm{L}}+T_y^{\mathrm{S}}$$

The cost rate c, turnover, and returns are expressed in percentage units; for example, $c=0.50$ denotes a cost rate of 0.50%. The year y refers to the portfolio formation year for the June-to-May holding period containing month m.

Therefore, the monthly long–short spread return, net of transaction costs, is defined as follows:

$$R_{p,m}^{†}\left(c\right)=R_{p,m}-{\displaystyle \frac{c\hspace{0.17em}\Delta w_y}{100\times 12}}$$

(8)

Figure 17b plots the arithmetic cumulative sum of monthly long–short spread returns after transaction costs. Over June 2013–May 2024, the terminal cumulative spread is 91.96% before costs and 69.27% under the 0.80% cost-rate scenario. The terminal cumulative spread remains positive under all scenarios examined, indicating that the combined strategy retains positive cumulative performance after the assumed trading costs.

5.4. Portfolio Composition by Industry and Firm Size

In portfolio management, concentration of portfolio constituents in a particular industry increases industry risk and limits diversification benefits. Smaller firms can also be more difficult to trade because of liquidity and market-impact considerations. Accordingly, for the long–short portfolio formed under the combined strategy, we examine the portfolio composition shares of selected stocks separately for the long and short legs.

Industry composition is calculated using the firms held in the portfolios. Firms are grouped using the TOPIX-17 classification, which distinguishes “Banks” from “Financials (ex Banks)”. The underlying firm-level information is available through Refinitiv Eikon.

Figure 18 reports the industry composition. On both the long and short legs, multiple industries, including “IT & Services, Others”, “Construction & Materials”, “Commercial & Wholesale Trade”, and “Electric Appliances & Precision Instruments”, account for meaningful shares, indicating diversification of industry risk.

Figure 18. Industry composition of portfolio holdings selected by the combined strategy (long–short portfolio).

Figure 19 reports the composition by firm size. Investable firms are sorted into market-capitalization quintiles (Q1–Q5), and the portfolio composition share of each quintile is shown. On the long leg, firms in the lower two market-capitalization quintiles (Q1–Q2) account for approximately half of the holdings, with the remainder drawn from Q3–Q5. The short leg also includes firms across the five size groups. Together, these results show that the combined strategy selects constituents across a broad range of firm sizes, with substantial representation of mid- and large-cap firms on both legs.

Figure 19. Same as Figure 18, but showing composition by firm size (market capitalization).

5.5. Discussion in Relation to the Research Questions

Regarding the first research question, multivariate XGBoost has a mean annual AUC of 0.86, compared with 0.77 for XGBoost using only the forecast dividend yield change. Its recall is 0.76, compared with 0.57 for the sign-based management forecast benchmark, although its precision is lower at 0.83 versus 0.95. As no formal statistical test of differences between classifiers is conducted, these comparisons should be interpreted descriptively. Within this limitation, the results indicate that combining firm-level information with dividend history information provides additional predictive information beyond management dividend forecasts.

Regarding the second research question, the combined strategy has an annualized Sharpe ratio of 1.26, compared with 0.71 and 0.44 for the yield-only and prediction-only strategies, respectively. Its three-factor alpha is positive and statistically significant (t = 4.12, p < 0.001). The alpha test evaluates whether the combined strategy’s alpha differs from zero and does not formally test whether its Sharpe ratio or alpha exceeds those of the other strategies. Subject to this limitation, the combined strategy shows the strongest overall performance among the three approaches under the evaluation settings used in this study, supporting the usefulness of combining forecast dividend yield with predicted dividend increase probability.

6. Conclusions

This study constructed machine-learning models to predict future dividend-increase behavior among firms listed on the TSE using financial indicators and other firm-level information and examined their effectiveness and potential applicability to asset management. Comparisons of XGBoost with other machine-learning models showed that dividend-increase behavior exhibited a degree of regularity and that the models achieved favorable predictive performance on metrics such as AUC. Because publicly disclosed dividend forecasts by Japanese firms are generally conservative, forecast-based predictions miss many firms that ultimately increase their dividends; the machine-learning models reduced such omissions and identified dividend-increasing firms with high recall. We also visualized feature importance and SHAP values for each explanatory variable to analyze the firm characteristics contributing to dividend-increase predictions and the direction of those contributions. The results indicate that forecast–actual dividend discrepancies, profitability, financial capacity, firm size, market valuation, ownership structure, and other characteristics contain important information for predicting dividend increases. These patterns are broadly consistent with existing theories of dividend rigidity, the corporate life cycle, financial capacity, and signaling.

To apply these findings to equity portfolio selection, we constructed a combined strategy using the predicted probability of a dividend increase and forecast dividend yield. Compared with the dividend-increase-prediction-only and traditional high-dividend-yield strategies, the combined strategy generated more stable long–short spread returns and a higher Sharpe ratio. Furthermore, a Fama–French three-factor regression showed that the combined strategy produced a statistically significant positive alpha after adjustment for conventional market, size, and value factors. After accounting for the transaction cost scenarios examined, the combined strategy retains a positive cumulative long–short spread return over the sample period. Its holdings span multiple industries and firm-size groups. In the compounded-cumulative-return analysis in Appendix D, the combined strategy also has the largest terminal CCRD between the top and bottom long-only portfolios. Together, these findings support the potential usefulness of combining dividend increase predictions with forecast dividend yield for stock selection.

Several limitations are directly related to the two research questions. For the first research question, the prediction results are based on a one-year rolling training window, which emphasizes recent relationships but does not establish that this window length is optimal. In addition, differences among the reported classification metrics are interpreted descriptively, and the feature importance and SHAP analyses identify predictive associations rather than causal mechanisms. Therefore, future research could examine alternative training windows, introduce a prespecified statistical framework for stricter model comparison, and investigate the economic mechanisms underlying the predictive relationships.

For the second research question, the portfolio results are conditional on the evaluation framework used in this study. The Fama–French three-factor model provides a conventional initial benchmark, but more advanced multifactor models and stricter portfolio evaluation settings would allow a broader assessment of performance robustness. In addition, the transaction cost analysis does not fully incorporate security-specific liquidity, borrowing conditions, short-sale feasibility, or realized execution costs. Therefore, future research should extend the risk-adjusted evaluation and implementation analysis. Generally, the findings remain conditional on the Japanese equity sample and evaluation period examined in this study.

Appendix A.1. Capital Efficiency Background

Figure A1. Trends in the percentage of firms listed on the TSE First Section before April 2022 and TSE Prime Market thereafter with return on equity below 8% and the percentage with a price-to-book ratio below 1.

Figure A1. Trends in the percentage of firms listed on the TSE First Section before April 2022 and TSE Prime Market thereafter with return on equity below 8% and the percentage with a price-to-book ratio below 1.

Appendix A.2. Variable-Level Relationships

For Figure A2 and Figure A3, continuous explanatory variables are divided into approximately 80 equal-width intervals after excluding observations outside the 1st–99th percentile range of each variable. Binary indicators are grouped by values of 0 and 1. Within each group, we calculate the mean variable value, observed proportion of dividend increases, and mean predicted probability. Similarly, model output is binned but without percentile trimming. These figures supplement the variable-level interpretation in Section 4.3.

Figure A2. Relationships between explanatory variable values and observed dividend increase rates and mean predicted probabilities. Note. The blue and red points indicate the observed incidence rates and mean model predictions within groups, respectively. For variables designated as “Standardized score” in Table 4, the horizontal axis shows the mean standardized value within each interval. Percentage signs in variable names refer to their original units before standardization.

Figure A2. Relationships between explanatory variable values and observed dividend increase rates and mean predicted probabilities. Note. The blue and red points indicate the observed incidence rates and mean model predictions within groups, respectively. For variables designated as “Standardized score” in Table 4, the horizontal axis shows the mean standardized value within each interval. Percentage signs in variable names refer to their original units before standardization.

Figure A3. Continuation of Figure A2 using the same grouping procedure and plotting conventions. Blue points indicate observed dividend-increase rates, and red points indicate mean predicted probabilities within each group.

Figure A3. Continuation of Figure A2 using the same grouping procedure and plotting conventions. Blue points indicate observed dividend-increase rates, and red points indicate mean predicted probabilities within each group.

Appendix B.1. Annual Sample Coverage

Table A1 distinguishes the base sample from the final machine learning sample for each annual dataset from 2011 to 2023. An observation refers to a firm in one annual dataset; the total is the sum of annual observations rather than the number of distinct firms over the full period. The year designation follows Section 4.1: information available at the end of May is paired with the dividend increase outcome over the following June-to-May period.

The final machine learning sample contains 13,165 observations, including 7195 dividend increases and 5970 nonincreases. The aggregate dividend increase rate is 54.65%, calculated from the total counts. For the 2013–2023 out-of-sample datasets, the final sample comprises 11,207 observations, including 6473 dividend increases.

Table A1. Annual sample counts, missing value exclusions, and dividend increase rates.

Year	Base N	Base Increase N	Missing Exclusions	Final ML N	Final Increase N	Final Nonincrease N	Final Rate [%]
2011	974	327	0	974	327	647	33.57
2012	984	395	0	984	395	589	40.14
2013	993	541	0	993	541	452	54.48
2014	1002	581	0	1002	581	421	57.98
2015	1013	608	0	1013	608	405	60.02
2016	1031	563	0	1031	563	468	54.61
2017	1089	657	0	1089	657	432	60.33
2018	1133	604	0	1133	604	529	53.31
2019	1166	489	0	1166	489	677	41.94
2020	541	248	1	540	247	293	45.74
2021	1135	774	1	1134	774	360	68.25
2022	1178	741	1	1177	740	437	62.87
2023	933	673	4	929	669	260	72.01
Total	13,172	7201	7	13,165	7195	5970	54.65

Note. The final dividend increase rate equals the final number of observations with a dividend increase divided by the final machine learning sample size. The total row rate is based on aggregate counts rather than the unweighted average of annual rates. Base and final samples refer to different processing stages.

Table A1. Annual sample counts, missing value exclusions, and dividend increase rates.

Year	Base N	Base Increase N	Missing Exclusions	Final ML N	Final Increase N	Final Nonincrease N	Final Rate [%]
2011	974	327	0	974	327	647	33.57
2012	984	395	0	984	395	589	40.14
2013	993	541	0	993	541	452	54.48
2014	1002	581	0	1002	581	421	57.98
2015	1013	608	0	1013	608	405	60.02
2016	1031	563	0	1031	563	468	54.61
2017	1089	657	0	1089	657	432	60.33
2018	1133	604	0	1133	604	529	53.31
2019	1166	489	0	1166	489	677	41.94
2020	541	248	1	540	247	293	45.74
2021	1135	774	1	1134	774	360	68.25
2022	1178	741	1	1177	740	437	62.87
2023	933	673	4	929	669	260	72.01
Total	13,172	7201	7	13,165	7195	5970	54.65

Note. The final dividend increase rate equals the final number of observations with a dividend increase divided by the final machine learning sample size. The total row rate is based on aggregate counts rather than the unweighted average of annual rates. Base and final samples refer to different processing stages.

Appendix B.2. Missing Value Treatment

After columns not used in the model were removed, records containing a missing value in any retained data column were excluded before continuous variable standardization. No missing input values were imputed at this step. The recorded transition from the base sample to the final machine learning sample excludes seven observations: one each in 2020, 2021, and 2022, and four in 2023. These exclusions account for approximately 0.0531% of the 13,172 base observations. These counts concern this model input preparation step, not any earlier preparation of the source datasets. Missing return exclusions in Section 3.3 are handled separately for the annual return calculations.

The substantial reduction in the 2020 sample size is largely attributable to the limited availability of dividend forecasts during the COVID-19 pandemic. According to the Tokyo Stock Exchange (2020), only 52.4% of the companies covered by the report disclosed dividend forecasts as of the end of May 2020, compared with 92.8% in the previous year. Since forecast-dividend information is required for our prediction model and portfolio construction, observations lacking such information at the prediction date were excluded, contributing to the smaller sample size in 2020. Consequently, the reduced sample size also limits the coverage of the corresponding model evaluation and portfolio analysis.

Appendix C.1. k-Nearest Neighbors

Table A4. Best hyperparameters for k-nearest neighbors by year.

Evaluation Year (Validation AUC)	n_neighbors	Weights	p	Algorithm	leaf_size
2013 (AUC = 0.7342)	20	distance	1	auto	43
2014 (AUC = 0.7831)	20	distance	1	auto	30
2015 (AUC = 0.8113)	20	distance	1	brute	40
2016 (AUC = 0.7876)	20	distance	1	brute	21
2017 (AUC = 0.7649)	19	distance	1	auto	49
2018 (AUC = 0.7885)	20	distance	1	auto	24
2019 (AUC = 0.7673)	20	distance	1	auto	49
2020 (AUC = 0.8120)	20	distance	1	brute	39
2021 (AUC = 0.7690)	20	distance	2	auto	41
2022 (AUC = 0.7095)	20	distance	1	brute	20
2023 (AUC = 0.7368)	15	distance	1	brute	31

Table A4. Best hyperparameters for k-nearest neighbors by year.

Evaluation Year (Validation AUC)	n_neighbors	Weights	p	Algorithm	leaf_size
2013 (AUC = 0.7342)	20	distance	1	auto	43
2014 (AUC = 0.7831)	20	distance	1	auto	30
2015 (AUC = 0.8113)	20	distance	1	brute	40
2016 (AUC = 0.7876)	20	distance	1	brute	21
2017 (AUC = 0.7649)	19	distance	1	auto	49
2018 (AUC = 0.7885)	20	distance	1	auto	24
2019 (AUC = 0.7673)	20	distance	1	auto	49
2020 (AUC = 0.8120)	20	distance	1	brute	39
2021 (AUC = 0.7690)	20	distance	2	auto	41
2022 (AUC = 0.7095)	20	distance	1	brute	20
2023 (AUC = 0.7368)	15	distance	1	brute	31

Appendix C.2. XGBoost

Table A5. Best hyperparameters for XGBoost by year (1).

Evaluation Year (Validation AUC)	learning_rate	n_estimators	max_depth	min_child_weight	Gamma
2013 (AUC = 0.8501)	0.0687	125	5	12	3.7070
2014 (AUC = 0.8587)	0.0342	137	4	7	2.8910
2015 (AUC = 0.8828)	0.0918	188	6	5	4.9900
2016 (AUC = 0.8908)	0.0507	159	3	8	1.1050
2017 (AUC = 0.8620)	0.0395	223	5	6	4.9260
2018 (AUC = 0.8694)	0.0509	153	3	15	3.0230
2019 (AUC = 0.8433)	0.0069	70	4	7	1.9950
2020 (AUC = 0.9026)	0.0804	115	3	15	4.7340
2021 (AUC = 0.8324)	0.0378	300	3	6	4.5510
2022 (AUC = 0.8648)	0.0255	258	6	5	3.1500
2023 (AUC = 0.8644)	0.0660	279	3	5	3.0440

Table A5. Best hyperparameters for XGBoost by year (1).

Evaluation Year (Validation AUC)	learning_rate	n_estimators	max_depth	min_child_weight	Gamma
2013 (AUC = 0.8501)	0.0687	125	5	12	3.7070
2014 (AUC = 0.8587)	0.0342	137	4	7	2.8910
2015 (AUC = 0.8828)	0.0918	188	6	5	4.9900
2016 (AUC = 0.8908)	0.0507	159	3	8	1.1050
2017 (AUC = 0.8620)	0.0395	223	5	6	4.9260
2018 (AUC = 0.8694)	0.0509	153	3	15	3.0230
2019 (AUC = 0.8433)	0.0069	70	4	7	1.9950
2020 (AUC = 0.9026)	0.0804	115	3	15	4.7340
2021 (AUC = 0.8324)	0.0378	300	3	6	4.5510
2022 (AUC = 0.8648)	0.0255	258	6	5	3.1500
2023 (AUC = 0.8644)	0.0660	279	3	5	3.0440

Table A6. Best hyperparameters for XGBoost by year (2).

Evaluation Year (Validation AUC)	Subsample	colsample_bytree	reg_alpha	reg_lambda	tree_method
2013 (AUC = 0.8501)	0.7438	0.9993	0.2171	1.2570	hist
2014 (AUC = 0.8587)	0.8123	0.8222	0.7149	1.6950	approx
2015 (AUC = 0.8828)	0.7288	0.8203	0.2819	1.2060	hist
2016 (AUC = 0.8908)	0.8600	0.7350	0.1811	1.4490	approx
2017 (AUC = 0.8620)	0.7210	0.7332	0.2331	0.7933	hist
2018 (AUC = 0.8694)	0.9649	0.7994	0.4541	0.9337	approx
2019 (AUC = 0.8433)	0.7063	0.8762	0.1593	1.3310	exact
2020 (AUC = 0.9026)	0.7594	0.9479	0.8656	1.1280	exact
2021 (AUC = 0.8324)	0.8272	0.7696	0.8742	0.9686	auto
2022 (AUC = 0.8648)	0.8179	0.8109	0.7225	1.8040	auto
2023 (AUC = 0.8644)	0.9077	0.7058	0.7519	1.6280	approx

Table A6. Best hyperparameters for XGBoost by year (2).

Evaluation Year (Validation AUC)	Subsample	colsample_bytree	reg_alpha	reg_lambda	tree_method
2013 (AUC = 0.8501)	0.7438	0.9993	0.2171	1.2570	hist
2014 (AUC = 0.8587)	0.8123	0.8222	0.7149	1.6950	approx
2015 (AUC = 0.8828)	0.7288	0.8203	0.2819	1.2060	hist
2016 (AUC = 0.8908)	0.8600	0.7350	0.1811	1.4490	approx
2017 (AUC = 0.8620)	0.7210	0.7332	0.2331	0.7933	hist
2018 (AUC = 0.8694)	0.9649	0.7994	0.4541	0.9337	approx
2019 (AUC = 0.8433)	0.7063	0.8762	0.1593	1.3310	exact
2020 (AUC = 0.9026)	0.7594	0.9479	0.8656	1.1280	exact
2021 (AUC = 0.8324)	0.8272	0.7696	0.8742	0.9686	auto
2022 (AUC = 0.8648)	0.8179	0.8109	0.7225	1.8040	auto
2023 (AUC = 0.8644)	0.9077	0.7058	0.7519	1.6280	approx

Appendix C.3. Random Forest

Table A7. Best hyperparameters for random forest by year (1).

Evaluation Year (Validation AUC)	n_estimators	max_depth	max_leaf_nodes	min_samples_split	min_samples_leaf
2013 (AUC = 0.8495)	85	4	792	3	6
2014 (AUC = 0.8596)	83	3	970	5	13
2015 (AUC = 0.8757)	53	7	770	3	12
2016 (AUC = 0.8853)	50	10	881	5	10
2017 (AUC = 0.8588)	160	4	771	7	5
2018 (AUC = 0.8662)	57	8	949	4	14
2019 (AUC = 0.8422)	187	8	793	6	15
2020 (AUC = 0.8968)	133	10	652	5	13
2021 (AUC = 0.8277)	63	8	700	3	15
2022 (AUC = 0.8554)	200	6	519	7	12
2023 (AUC = 0.8577)	64	10	571	3	13

Note. AUC, area under the receiver operating characteristic curve.

Table A7. Best hyperparameters for random forest by year (1).

Evaluation Year (Validation AUC)	n_estimators	max_depth	max_leaf_nodes	min_samples_split	min_samples_leaf
2013 (AUC = 0.8495)	85	4	792	3	6
2014 (AUC = 0.8596)	83	3	970	5	13
2015 (AUC = 0.8757)	53	7	770	3	12
2016 (AUC = 0.8853)	50	10	881	5	10
2017 (AUC = 0.8588)	160	4	771	7	5
2018 (AUC = 0.8662)	57	8	949	4	14
2019 (AUC = 0.8422)	187	8	793	6	15
2020 (AUC = 0.8968)	133	10	652	5	13
2021 (AUC = 0.8277)	63	8	700	3	15
2022 (AUC = 0.8554)	200	6	519	7	12
2023 (AUC = 0.8577)	64	10	571	3	13

Note. AUC, area under the receiver operating characteristic curve.

Table A8. Best hyperparameters for random forest by year (2).

Evaluation Year (Validation AUC)	max_features	class_weight
2013 (AUC = 0.8495)	0.9918	balanced
2014 (AUC = 0.8596)	0.7713	balanced
2015 (AUC = 0.8757)	0.7035	None
2016 (AUC = 0.8853)	0.7472	balanced
2017 (AUC = 0.8588)	0.8673	None
2018 (AUC = 0.8662)	0.8570	balanced
2019 (AUC = 0.8422)	0.9490	balanced
2020 (AUC = 0.8968)	0.7448	None
2021 (AUC = 0.8277)	0.7123	None
2022 (AUC = 0.8554)	0.9764	None
2023 (AUC = 0.8577)	0.7414	balanced

Table A8. Best hyperparameters for random forest by year (2).

Evaluation Year (Validation AUC)	max_features	class_weight
2013 (AUC = 0.8495)	0.9918	balanced
2014 (AUC = 0.8596)	0.7713	balanced
2015 (AUC = 0.8757)	0.7035	None
2016 (AUC = 0.8853)	0.7472	balanced
2017 (AUC = 0.8588)	0.8673	None
2018 (AUC = 0.8662)	0.8570	balanced
2019 (AUC = 0.8422)	0.9490	balanced
2020 (AUC = 0.8968)	0.7448	None
2021 (AUC = 0.8277)	0.7123	None
2022 (AUC = 0.8554)	0.9764	None
2023 (AUC = 0.8577)	0.7414	balanced

Appendix C.4. Logistic Regression

Table A9. Best hyperparameters for logistic regression by year.

Evaluation Year (Validation AUC)	C	Penalty	class_weight	max_iter	tol
2013 (AUC = 0.8296)	0.1006	l1	balanced	1208	0.0018
2014 (AUC = 0.8457)	0.3423	l1	None	1612	0.0009
2015 (AUC = 0.8567)	0.2499	l1	balanced	1934	0.0005
2016 (AUC = 0.8663)	0.6191	l1	balanced	729	0.0009
2017 (AUC = 0.8463)	0.0950	l1	balanced	911	0.0006
2018 (AUC = 0.8376)	0.1736	l1	balanced	1162	0.0049
2019 (AUC = 0.8121)	0.1839	l1	balanced	1020	0.0014
2020 (AUC = 0.8706)	0.3176	l1	None	1989	0.0006
2021 (AUC = 0.8204)	0.0433	l2	balanced	1881	0.0005
2022 (AUC = 0.8194)	0.2448	l1	balanced	1329	0.0003
2023 (AUC = 0.8232)	0.1475	l1	balanced	1943	0.0022

Table A9. Best hyperparameters for logistic regression by year.

Evaluation Year (Validation AUC)	C	Penalty	class_weight	max_iter	tol
2013 (AUC = 0.8296)	0.1006	l1	balanced	1208	0.0018
2014 (AUC = 0.8457)	0.3423	l1	None	1612	0.0009
2015 (AUC = 0.8567)	0.2499	l1	balanced	1934	0.0005
2016 (AUC = 0.8663)	0.6191	l1	balanced	729	0.0009
2017 (AUC = 0.8463)	0.0950	l1	balanced	911	0.0006
2018 (AUC = 0.8376)	0.1736	l1	balanced	1162	0.0049
2019 (AUC = 0.8121)	0.1839	l1	balanced	1020	0.0014
2020 (AUC = 0.8706)	0.3176	l1	None	1989	0.0006
2021 (AUC = 0.8204)	0.0433	l2	balanced	1881	0.0005
2022 (AUC = 0.8194)	0.2448	l1	balanced	1329	0.0003
2023 (AUC = 0.8232)	0.1475	l1	balanced	1943	0.0022

Appendix C.5. SVM

Table A10. Best hyperparameters for SVM by year (1).

Evaluation Year (Validation AUC)	C	Kernel	Gamma	class_weight	tol
2013 (AUC = 0.8223)	0.0086	linear	scale	None	0.0026
2014 (AUC = 0.8446)	995.0000	linear	scale	None	0.0004
2015 (AUC = 0.8529)	0.7276	linear	scale	balanced	0.0095
2016 (AUC = 0.8638)	9.5190	linear	scale	balanced	0.0017
2017 (AUC = 0.8383)	0.0113	linear	scale	None	0.0080
2018 (AUC = 0.8308)	0.0608	linear	scale	None	0.0065
2019 (AUC = 0.8127)	0.1149	linear	scale	balanced	0.0052
2020 (AUC = 0.8692)	0.2603	linear	scale	balanced	0.0036
2021 (AUC = 0.8156)	0.0179	linear	auto	None	0.0072
2022 (AUC = 0.8066)	0.0155	linear	auto	None	0.0048
2023 (AUC = 0.8171)	2.0350	linear	auto	balanced	0.0051

Table A10. Best hyperparameters for SVM by year (1).

Evaluation Year (Validation AUC)	C	Kernel	Gamma	class_weight	tol
2013 (AUC = 0.8223)	0.0086	linear	scale	None	0.0026
2014 (AUC = 0.8446)	995.0000	linear	scale	None	0.0004
2015 (AUC = 0.8529)	0.7276	linear	scale	balanced	0.0095
2016 (AUC = 0.8638)	9.5190	linear	scale	balanced	0.0017
2017 (AUC = 0.8383)	0.0113	linear	scale	None	0.0080
2018 (AUC = 0.8308)	0.0608	linear	scale	None	0.0065
2019 (AUC = 0.8127)	0.1149	linear	scale	balanced	0.0052
2020 (AUC = 0.8692)	0.2603	linear	scale	balanced	0.0036
2021 (AUC = 0.8156)	0.0179	linear	auto	None	0.0072
2022 (AUC = 0.8066)	0.0155	linear	auto	None	0.0048
2023 (AUC = 0.8171)	2.0350	linear	auto	balanced	0.0051

Table A11. Best hyperparameters for SVM by year (2).

Evaluation Year (Validation AUC)	Degree	coef0	Shrinking
2013 (AUC = 0.8223)	Not applicable	Not applicable	False
2014 (AUC = 0.8446)	Not applicable	Not applicable	True
2015 (AUC = 0.8529)	Not applicable	Not applicable	True
2016 (AUC = 0.8638)	Not applicable	Not applicable	True
2017 (AUC = 0.8383)	Not applicable	Not applicable	True
2018 (AUC = 0.8308)	Not applicable	Not applicable	False
2019 (AUC = 0.8127)	Not applicable	Not applicable	True
2020 (AUC = 0.8692)	Not applicable	Not applicable	True
2021 (AUC = 0.8156)	Not applicable	Not applicable	True
2022 (AUC = 0.8066)	Not applicable	Not applicable	False
2023 (AUC = 0.8171)	Not applicable	Not applicable	False

Note: 1. degree was included in the search only when kernel = poly; otherwise, it was treated as not applicable. 2. coef0 was included in the search only when kernel = poly or kernel = sigmoid; otherwise, it was treated as not applicable. 3. probability was fixed at False, and AUC was calculated using the output of decision_function.

Table A11. Best hyperparameters for SVM by year (2).

Evaluation Year (Validation AUC)	Degree	coef0	Shrinking
2013 (AUC = 0.8223)	Not applicable	Not applicable	False
2014 (AUC = 0.8446)	Not applicable	Not applicable	True
2015 (AUC = 0.8529)	Not applicable	Not applicable	True
2016 (AUC = 0.8638)	Not applicable	Not applicable	True
2017 (AUC = 0.8383)	Not applicable	Not applicable	True
2018 (AUC = 0.8308)	Not applicable	Not applicable	False
2019 (AUC = 0.8127)	Not applicable	Not applicable	True
2020 (AUC = 0.8692)	Not applicable	Not applicable	True
2021 (AUC = 0.8156)	Not applicable	Not applicable	True
2022 (AUC = 0.8066)	Not applicable	Not applicable	False
2023 (AUC = 0.8171)	Not applicable	Not applicable	False

Note: 1. degree was included in the search only when kernel = poly; otherwise, it was treated as not applicable. 2. coef0 was included in the search only when kernel = poly or kernel = sigmoid; otherwise, it was treated as not applicable. 3. probability was fixed at False, and AUC was calculated using the output of decision_function.

Appendix C.6. Software and Implementation

The analysis and visualization code was written in Python 3.11. pandas and NumPy were used for tabular data processing and numerical operations, scikit-learn for preprocessing and classification metrics, XGBoost for boosted-tree models, and SHAP for model interpretation. SciPy was used for t-distribution calculations. The three-factor regression code estimates ordinary least squares coefficients with NumPy and implements the Newey–West covariance calculation in a custom function. Figures were produced with Matplotlib (version 3.10.8) and seaborn. The Optuna hyperparameter search procedure is described in Section 4.1; stored search results are read for subsequent model fitting.

Appendix D.1. Compounded Return Calculation Under Annual Rebalancing

Let p denote a portfolio, y the portfolio formation year, and m = 1, …, 12 the months from June of year y through May of year y + 1. Let $C_{p,y}$ denote the capital available at the beginning of formation year y, and let $N_{p,y}$ denote the number of stocks selected for portfolio p.

At portfolio formation, the available capital is allocated equally across the selected stocks. Therefore, the initial value assigned to stock i is as follows:

$$v_{i,p,y,0}={\displaystyle \frac{C_{p,y}}{N_{p,y}}}$$

(A1)

Let $r_{i,p,y,m}$ denote the monthly return of stock i, expressed in decimal form, using the same return inputs as in Section 5.1. During the 12-month holding period, no within-year rebalancing is performed. Instead, the value of each stock position evolves according to

$$v_{i,p,y,m}=v_{i,p,y,m-1}\left(1+r_{i,p,y,m}\right).$$

(A2)

Then, the month-end value of portfolio p is obtained by summing the values of all constituent positions:

$$V_{p,y,m}={\displaystyle {\sum }_{i=1}^{N_{p,y}}}{v}_{i,p,y,m}$$

(A3)

As the individual positions evolve according to their own realized returns, portfolio weights are allowed to drift naturally within the holding year. Equal weights are restored only at the next annual portfolio formation.

At the end of May, the closing portfolio value becomes the capital available for the following formation year:

$$C_{p,y+1}=V_{p,y,12}$$

(A4)

Thus, gains and losses are reflected continuously in portfolio value, while portfolio constituents and equal target weights are updated only once per year.

For the small number of missing monthly return observations in the selected stock matrices, the position value from the preceding month is carried forward for that month. This treatment preserves the annual holding rule and avoids mechanically reallocating capital among the remaining stocks when return availability changes.

The compounded cumulative return of portfolio p at month t is defined relative to its initial capital as

$$C{R}_{p,t}=100\left({\displaystyle \frac{V_{p,t}}{V_{p,0}}}-1\right)$$

(A5)

All portfolios begin with the same normalized initial capital; therefore, their compounded cumulative return paths can be directly compared.

Appendix D.2. Compounded Cumulative Return Differential (CCRD) and Benchmark Presentation

To further evaluate the stock-selection performance under compounded returns, we introduce the Compounded Cumulative Return Differential (CCRD) as a supplementary measure. This measure examines whether the performance differentiation between the top and bottom stock groups observed in the main analysis remains evident when returns are compounded over time.

TOPIX is shown as a separate market benchmark rather than being subtracted from each portfolio return. Using the same monthly TOPIX return series as in the main analysis, its normalized value evolves as

$$V_{\mathrm{M},t}=V_{\mathrm{M},t-1}\left(1+r_{\mathrm{M},t}\right), V_{\mathrm{M},0}=1$$

(A6)

and its compounded cumulative return is 100($V_{\mathrm{M},t}$ − 1).

For a descriptive comparison with the top-minus-bottom results in Section 5.1, CCRD is calculated as the difference between the compounded cumulative returns of the corresponding upper and lower portfolios, each formed and compounded independently as a long-only portfolio. Let $G_{s,t}^{\mathrm{T}}$ and $G_{s,t}^{\mathrm{B}}$ denote the normalized portfolio value factors of the top and bottom long-only portfolios for strategy s, respectively. The CCRD, expressed in percentage points, is defined as follows:

$${\mathrm{C}\mathrm{C}\mathrm{R}\mathrm{D}}_{s,t}=C{R}_{s,t}^{\mathrm{T}}-C{R}_{s,t}^{\mathrm{B}}=100\left(G_{s,t}^{\mathrm{T}}-G_{s,t}^{\mathrm{B}}\right)$$

(A7)

Both portfolios are financed and compounded separately as long-only portfolios. Accordingly, CCRD is a descriptive comparison and should not be interpreted as the cumulative wealth or compounded return of a self-financing long–short strategy.

Figure A4 and Figure A5 report the resulting compounded cumulative returns. Transaction costs are not deducted in these figures because the purpose of this appendix is specifically to examine the effect of compounding; turnover and transaction cost effects are evaluated separately in Section 5.3.

Appendix D.3. Results and Implications for Stock Selection

Figure A4 applies the compounded return calculation to the portfolios formed on predicted dividend increase probabilities in Figure 14.

Figure A4. Compounded cumulative returns of portfolios formed on predicted dividend increase probabilities. Note. The lines show compounded cumulative returns for the top, middle, and bottom predicted-probability terciles, perfect foresight, and TOPIX. The pink bars indicate CCRD (Compounded Cumulative Return Differential), defined as the difference between the compounded cumulative returns of the separately financed top and bottom long-only portfolios, expressed in percentage points. Perfect foresight is an ex post, nontradable reference.

Figure A4. Compounded cumulative returns of portfolios formed on predicted dividend increase probabilities. Note. The lines show compounded cumulative returns for the top, middle, and bottom predicted-probability terciles, perfect foresight, and TOPIX. The pink bars indicate CCRD (Compounded Cumulative Return Differential), defined as the difference between the compounded cumulative returns of the separately financed top and bottom long-only portfolios, expressed in percentage points. Perfect foresight is an ex post, nontradable reference.

The probability-sorted portfolios retain the broad performance ordering observed in Figure 14. At the end of the evaluation period, the compounded cumulative returns are 225.58% for perfect foresight, 146.53% for the top predicted-probability tercile, 125.38% for the middle tercile, and 93.88% for the bottom tercile. TOPIX records a compounded cumulative return of 104.73%. Thus, the top portfolio finishes above the market benchmark, whereas the bottom portfolio finishes below it. The terminal CCRD between the top and bottom long-only portfolios is 52.65 percentage points.

These results indicate that the predicted probability of a dividend increase continues to distinguish stocks with stronger and weaker subsequent performance when returns are compounded through the holding period. Therefore, the robustness check supports the stock selection interpretation of the prediction signal reported in Section 5.1.

Figure A5 applies the same calculation to the three investment strategies compared in Figure 16.

Figure A5. Comparison of compounded cumulative portfolio performance across the three investment strategies. Note. Panels (a–c) show the compounded cumulative returns of the corresponding upper and lower stock groups together with TOPIX. The pink bars show CCRD, the compounded cumulative return differential between separately financed top and bottom long-only portfolios. Panel (d) compares CCRD across the three strategies on the same scale.

Figure A5. Comparison of compounded cumulative portfolio performance across the three investment strategies. Note. Panels (a–c) show the compounded cumulative returns of the corresponding upper and lower stock groups together with TOPIX. The pink bars show CCRD, the compounded cumulative return differential between separately financed top and bottom long-only portfolios. Panel (d) compares CCRD across the three strategies on the same scale.

The terminal CCRD values are 106.64, 52.65, and 197.85 percentage points for the forecast-dividend-yield-only, dividend-increase-prediction-only, and combined strategies, respectively. Therefore, the combined strategy exhibits the largest CCRD among the three approaches. These values describe differences between independent long-only portfolios and are not compounded returns from self-financing long–short investment accounts.

This ordering is consistent with the main results in Figure 16. Therefore, the additional compounded return analysis does not alter the principal stock-selection conclusion: combining the predicted probability of a dividend increase with forecast dividend yield provides stronger stock selection performance than using either signal alone. Moreover, the result supports the interpretation that forecast dividend yield and predicted dividend increase probability provide complementary information on the current level of dividend income and prospect of future dividend growth.