MDMF: A Market-Mainline-Driven Multi-Feature Fusion Model for Stock Trend Forecasting

Abstract

Accurate stock prediction is crucial for investment decisions and risk management, yet remains challenging due to the non-stationary, nonlinear, and noisy nature of financial markets. Although deep learning has advanced forecasting by modeling temporal patterns and stock relationships, most methods fail to capture structured, market-wide forces. Specifically, they miss the emergence and influence of “market mainlines”—persistent directional trends in groups of stocks with shared attributes that collectively drive market movement. To address this, we propose the Market-mainline-Driven Multi-feature Fusion Model (MDMF), which dynamically identifies multidimensional market mainline characteristics and captures their differential impact on individual stocks. The model incorporates a dual-channel encoding mechanism, a dynamic stock aggregation algorithm, and a differential influence module to integrate temporal, fundamental, and stock-specific features. Extensive experiments on real-world stock datasets show that MDMF outperforms state-of-the-art baselines in predictive accuracy and profitability, demonstrating its robustness and practical utility. Our study highlights the value of explicitly modeling market mainlines for enhancing stock prediction and offers insights into systematic market behavior.

1. Introduction

The application of machine learning in stock prediction has consistently attracted significant attention as a core research topic in financial technology. High-quality stock price forecasting not only assists investors in formulating strategies but also holds substantial practical significance for risk management and asset allocation. However, the stock market is subject to compound effects from macroeconomic fluctuations, industrial policies, and investor behaviors. This results in a system characterized by profound complexity, manifesting as high non-stationarity, strong nonlinearity, and a low signal-to-noise ratio. These attributes make trend analysis and price prediction in stock markets exceptionally challenging.

In recent years, deep learning has greatly improved stock prediction by effectively analyzing historical prices, trading volumes, and event-based news [1,2,3,4]. Many studies focus on predicting individual stocks using models like RNNs and Transformers to identify patterns in complex, time-series data, leading to better forecasting results [5,6,7]. To better capture complex relationships between stocks, recent research integrates market-logic-based networks into prediction models. These networks use the influence of one stock on another as valuable data to enhance accuracy [8,9,10]. Although these recent studies have revealed the potential of mining inter-stock relationships for enhancing stock trend prediction, they still exhibit certain limitations. Furthermore, stock prices are influenced not only by other individual stocks but also by broader market dynamics, such as industry performance, investor sentiment, and macroeconomic indicators. Effectively leveraging and learning from these integrated market dynamics remains an open challenge.

Recent studies have begun exploring ways to leverage valuable information embedded in overall market dynamics to enhance stock trend forecasting. For instance, Chen et al. [11] utilize fundamental stock data to construct market dynamic features and mine correlations between individual stocks and market characteristics to improve predictive performance. Additionally, Chen et al. [12] attempts to extract key features from market performance and macroeconomic data to strengthen the model’s understanding of complex market mechanisms. However, existing models of market dynamics often rely solely on macro-level indicators, failing to break down internal market forces from a structural or behavioral perspective. These approaches largely overlook the structural mainlines present in the market and their differential pulling effects on individual stocks, thereby limiting the expressiveness and interpretability of prediction models. Therefore, to better capture meaningful signals from overall market movements, it is essential to identify key market drivers behind upward trends and analyze their complex relationships with individual stocks.

A common phenomenon in the stock market is that during certain periods, groups of stocks with similar attributes—such as industry, theme, market cap, or risk profile—often show strong price trends. These movements not only exceed broader market performance but can also drive its overall direction. This systematic trend, led by specific clusters of stocks, is known as the market mainline. The formation and evolution of the market mainline are primarily driven by two herding mechanisms: spurious herding and intentional herding [13]. Spurious herding occurs when investors independently yet convergently react to common fundamental shocks. In contrast, intentional herding arises when investors consciously adjust their strategies after observing the actions of others. Specifically, the market mainline often initially takes shape due to spurious herding triggered by external fundamental signals, or emerges through intentional herding as some investors mimic the strategies of others. Once the mainline trend is established and gains widespread market attention, its influence reinforces itself through bandwagon effects, continuously attracting capital inflows and consolidating its dominant market position [14]. More importantly, the market mainline does not affect all stocks uniformly. Instead, it exerts a sustained and differentiated driving force on individual stocks depending on the alignment between their attributes and the mainline’s underlying logic. This results in a pronounced structural characteristic across the market.

For instance, from August 2020 to early 2021, amid expectations of accommodative macro liquidity and substantial public fund issuance, a structural trend centered on “core assets” emerged as the market mainline. Capital consistently flowed into large-cap blue-chip stocks, lifting the CSI 300 index markedly, while the CSI 500 and CSI 1000 indices underperformed relatively (Figure 1a). This illustrates the formation and strengthening of a market mainline: driven by liquidity and institutional allocation, the large-cap blue-chip sector attracted sustained capital inflows, and its growing return advantage steered the overall market style. Later, by the end of 2021, supported by the advancing dual-carbon policy and strong sector fundamentals, the new energy sector converged into a new market mainline, drawing intensive capital. In contrast, the traditional energy sector lagged significantly due to its misalignment with the prevailing industrial logic (Figure 1b). These two market episodes illustrate that once a market mainline is established, it continuously attracts capital concentration through herding mechanisms and expectation reinforcement, exerting a differentiated pulling effect on individual stocks. Stocks aligned with the mainline logic persistently garner capital favor, whereas those with misaligned attributes may suffer significant relative losses due to capital outflow, even during phases of overall market appreciation.

Therefore, to better capture how the market trend and individual stocks interact, and to give decision-makers a clear view of market dynamics, we propose the Market-mainline-Driven Multi-feature Fusion Model (MDMF). The model dynamically identifies the multidimensional features of the market mainline and precisely captures its varying effects on different stocks, improving the accuracy of stock trend predictions. In its design, the MDMF emphasizes the following three aspects:

(1)

The formation of the market mainline involves complex and volatile factors—such as price fluctuations, economic cycles, and policy directions. To better capture its dynamic evolution, we propose a dual-channel stock feature encoding method combined with a dynamic stock set aggregation algorithm. Specifically, a multidimensional representation of stock features is collaboratively constructed through both temporal and fundamental channels. Meanwhile, the dynamic stock set aggregation algorithm extracts latent market mainline features from specific groups of stocks. The representation of these features is then refined by uncovering the underlying relationships between all stocks and the market mainline.

(2)

To deeply explore the underlying connections between individual stocks and the multidimensional dynamic market state, a differential impact mechanism of the market mainline is further introduced. This mechanism captures the complex relationships between the market mainline and individual stocks from multiple perspectives. Specifically, within the temporal and fundamental dimensions, two distinct algorithms are employed to dynamically analyze the differentiated influence of the market mainline on various stocks.

(3)

To effectively extract the individual characteristics of each stock, the impact of the dynamic market mainline is filtered out to isolate stock-specific attributes. On this basis, multidimensional influence information from the market mainline is integrated with idiosyncratic stock features to achieve accurate prediction of stock price trends.

We conducted comprehensive evaluations of the proposed framework on real-world stock datasets. Experimental results demonstrate that our model achieves superior performance across all evaluation metrics compared to multiple baseline methods. Furthermore, we implemented an industry-standard trading strategy to simulate stock investment processes, with results showing our framework delivers higher cumulative returns than all baselines. We also performed additional analyses to investigate the effects of different components within our framework, further revealing the advantages of MDMF.

Contributions of this work include the following:

(1)

We proposes a novel model that captures dynamic evolutionary characteristics of market mainlines from both temporal and fundamental perspectives, effectively addressing the challenge of complex and volatile formation factors.

(2)

Our model enhances stock trend prediction performance by simultaneously leveraging differentiated impacts of temporal and fundamental market mainlines on individual stocks alongside stock-specific attributes.

(3)

We validated the effectiveness of our MDMF framework through experimental evaluations and investment simulations on real-world stock market datasets.

2. Related Works

Research on predicting future stock price trends in recent years can generally be categorized into two major classes: single-stock based methods and cross-stock based methods. The former method uses only historical data from individual stocks to identify patterns within their own price sequences. In contrast, the latter also incorporates complex, market-driven relational networks among stocks. By integrating these cross-stock influences, the model captures a wider range of market co-movement patterns.

2.1. Single-Stock-Based Methods

Among single-stock based approaches, traditional statistical forecasting models such as Autoregressive (AR) [15] and ARIMA [16] are suitable for linear and stationary time series, and have been widely adopted in this domain. However, the nonlinear and chaotic nature of stock prices limits the applicability of these statistical models. Given the limitations of statistical models in handling nonlinear data, several studies have attempted to apply deep neural networks to capture complex patterns in market trends [17,18]. To model long-term dependencies in time series, Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) [19] networks has been extensively utilized [20,21,22,23]. Specifically, Burak [24] introduced the Artificial Rabbits Optimization algorithm to automatically optimize the hyperparameters of LSTM; Ali et al. [25] enhanced the decomposition of trend and fluctuation components in financial time series by introducing an improved Empirical Mode Decomposition (EMD) method that incorporates Akima spline interpolation for preprocessing raw data; Lu and Xu [26] constructed a TRNN model that integrates time-series compression techniques with a dynamic price–volume relationship to boost training efficiency and amplify the impact of recent trading volume on stock prices. Meanwhile, Transformer-based models have also achieved remarkable progress due to their powerful long-sequence modeling capabilities and self-attention mechanisms [27,28,29]. Specifically, Zhang et al. [30] developed the TEANet framework by integrating Transformer encoder and multi-attention mechanisms to achieve multimodal fusion of social media text and stock prices, aiming to address the temporal dependency issues in financial data and enhance prediction accuracy under small-sample conditions; Li et al. [31] introduced a market-guided mechanism and an alternating aggregation strategy to model instantaneous and cross-time dynamic stock correlations and achieve adaptive feature selection; and Muhammad et al. [32] integrated the time2vec time encoding technique and applied the Transformer model to Dhaka Stock Exchange (DSE) data for the first time, aiming to validate the model’s predictive capability for volatile stock prices in emerging markets.

However, individual stocks do not evolve independently. Methods focusing solely on single stocks struggle to capture complex inter-stock relationships, which constrains the model’s ability to understand and predict stock behavior.

2.2. Cross-Stock-Based Methods

To mine cross-stock relational information and enhance stock trend prediction performance, recent cross-stock methods have employed graph modeling frameworks to capture relationships among different stocks [33,34,35]. Specifically, Xiang et al. [36] achieved precise modeling of dynamic stock relationships by constructing temporally dynamic graphs generated from daily pricing data, combined with Transformer encoding and heterogeneous graph attention networks. Ma et al. [37] automatically generates multidimensional stock correlation graphs from historical data to replace predetermined graphs, aiming to achieve in-depth exploration of hidden inter-stock correlations and temporal dependencies. Qian et al. [38] developed a discrete dynamic graph structure integrated with Transformer temporal encoding, aiming to capture the dynamic and complex inter-stock relationships arising from economic indicators, financial reports as well as their temporal evolution patterns. Han et al. [39] employed a hypergraph structure to characterize high-order industry relationships and dynamically update stock correlations, effectively addressing the limitation of traditional models in capturing industry-level high-order interactions and significantly enhancing the accuracy of stock trend prediction.

While such approaches model stock relationships via predefined structures and aggregate information through graph/hypergraph networks, their static architectures fail to capture dynamic market drivers like policy changes or capital flow variations. Thus, despite reflecting group characteristics, they neglect critical individual- and sector-level market linkages. Thus, despite reflecting group characteristics, they neglect correlations between individual stocks, industries, and broader market dynamics. Aiming to capture significant fluctuations triggered within stock groups sharing common characteristics (e.g., industry, region, concept, or similar volatility), Shi et al. [40] constructed four types of graph structures incorporating different relational knowledge, utilizing Graph Convolutional Network (GCN) to extract multi-period stock embedding features. Liu et al. [41] enhanced the representational efficacy of TCN for multidimensional price sequences and achieved more accurate stock ranking predictions by incorporating dynamic correlations between industry attributes and market preferences. Liu et al. [42] utilized dynamic spatio-temporal graph convolutional layers to capture complex dynamic stock correlations across industries, companies, and time dimensions, significantly enhancing the accuracy of stock price predictions. Wang et al. [43] developed a hybrid architecture that integrates the semantic relationships of knowledge graphs with the topological reasoning capabilities of graph convolutional networks, quantifying both explicit and implicit inter-stock correlations into computable topological features. By combining community detection algorithms to identify similar stock clusters, the approach ultimately establishes a collaborative prediction mechanism among stock communities to overcome the limitations of traditional models in capturing deep underlying factors.

However, existing research on modeling market states often focuses on the overall market or relies on fixed industry classifications. Consequently, these approaches aggregate information either using all stocks or through industry relationships, neglecting dynamically evolving relational structures within the market. Furthermore, market features are often constructed from a single static dimension (such as macroeconomic indicators), making it difficult to capture the multidimensional interactions between individual stocks and complex market states.

3. Methodology

Figure 2 illustrates MDMF’s overall architecture through three sequential processing stages. Initially, multidimensional features are extracted for each stock using a stock feature encoder, while a market mainline feature encoder is constructed via graph neural networks to capture dynamically evolving mainline information. Subsequently, the obtained multidimensional stock features are integrated with an attention mechanism to mine differentiated effects of multiple market mainlines across individual stocks, concurrently processing stock-independent characteristics through a dedicated module. Ultimately, the extracted information is processed by a feedforward network. This network combines the multidimensional effects from market mainlines with trends unique to individual stocks, leading to more accurate predictions of stock price movements. The stock trend prediction problem is formally defined in this section alongside detailed methodology elaboration.

3.1. Problem Definition

Given stock-specific features (e.g., historical prices, volumes, fund manager holdings) for stock i at date t, stock trend prediction aims to forecast the price trend $d_i^t$, defined as the next-day price change rate:

$$d_i^t={\displaystyle \frac{{\mathrm{Price}}_i^{t+1}-{\mathrm{Price}}_i^t}{{\mathrm{Price}}_i^t}}$$

(1)

The Price metric may be specified using different values, such as the opening price, closing price, or volume-weighted average price (VWAP). In this study, the closing price is employed to compute stock price trends.

3.2. Stock Feature Encoding

The formation of market mainlines in financial markets stems from complex and diverse drivers that manifest distinct characteristics across different time periods. These mainlines may emerge from capital-flow-induced shifts in stock price–volume dynamics, collective sector fluctuations triggered by economic cycles, policy directives, or other catalysts. For instance, during the COVID-19 pandemic in 2020, surging remote-work demand drove substantial price surges in technology and communication stocks. This mainline formation primarily originated from short-term market sentiment and price–volume variations—essentially temporal features. Conversely, in 2022, China’s “dual-carbon” policy advancement established carbon neutrality as a dominant market mainline, demonstrating policy-driven mainline evolution underpinned by corporate fundamental attributes. To comprehensively capture the intricate relationships between market mainlines and individual stocks, we propose a dual-channel stock feature encoding method. This method constructs multidimensional stock representations through the synergistic integration of temporal and fundamental feature channels.

3.2.1. Temporal Feature Extraction

Stock price–volume time-series data represents a canonical high-noise, non-stationary sequence whose dynamic patterns are embedded within complex long-range dependencies. Yet not every historical observation is equally informative for the future: market noise and short-term fluctuations require filtration, while historically significant states pertinent to current market conditions must be preserved and propagated. To address this selective memory and dynamic update requirement, we analyze temporal patterns within price–volume sequences to capture stock price dynamics.

Specifically, given input features ${\mathcal{S}}_t=\left\{s_1^t,s_2^t,\dots ,s_n^t\right\}$ for n stocks at date t, where $s_i^t=\left\{x_i^{t-T-1},\dots ,x_i^t\right\}\in {\mathbb{R}}^{T\times d}$ represents the price–volume time-series of stock i over T preceding trading days. Given the coexistence of critical and secondary information in stock price–volume data, we introduce a forget mechanism $r_i^t=\mathrm{sigmoid}(W_r[p_i^{t-1},s_i^t]+b_r)$ to filter out less relevant historical information pertaining to current price dynamics. This mechanism learns to filter historically irrelevant information based on the prior state $p_i^{t-1}$ and current input $s_i^t$, with $W_r$ and $b_r$ denoting trainable weight matrix and bias vector. Simultaneously accounting for the differential impact of historical price–volume data on new states, we introduce an update mechanism $z_i^t=\mathrm{sigmoid}(W_z[p_i^{t-1},s_i^t]+b_z)$ to regulate information retention from previous states and integration into current representations, thereby efficiently capturing stock price dynamics. Here, $W_z$ and $b_z$ denote trainable weight matrix and bias vector respectively.

Building upon these mechanisms, we establish the following method to capture long-term dependencies and learn temporal stock features, where the final hidden state $p_i^t\in {\mathbb{R}}^h$ (h denotes the hidden dimension) serves as the initial embedding for stock i at date t:

$$p_i^t=(1-z_i^t)\odot p_i^{t-1}+z_i^t\odot tanh\left(W_h\left[r_i^t\odot p_i^{t-1},s_i^t\right]+b_h\right)$$

(2)

where ⊙ denotes the element-wise multiplication operator, and $W_h$ and $b_h$ are trainable parameters.

The initial embeddings of all stocks at day t form the matrix $P_t={[p_1^t,\dots ,p_n^t]}^{\top }\in {\mathbb{R}}^{n\times h}$, which provides input representations for downstream tasks.

3.2.2. Fundamental Feature Extraction

Mutual fund portfolio reports, released semi-annually in regulated markets such as China, constitute structured records of fund managers’ investment behaviors that reflect their strategic positioning across thousands of stocks. These actions intrinsically encode managers’ expertise and focus on stocks’ fundamental attributes—latent properties driving investment decisions, including sector alignment (e.g., financials, information technology), economic cycle sensitivity (e.g., cyclical versus defensive stocks), or other abstract financial characteristics [11]. Through analyzing fund managers’ portfolios, we capture their persistent preferences for specific stocks, thereby revealing fundamental stock features.

Specifically, we construct a fund-stock holding tensor $\mathcal{M}\in {\mathbb{R}}^{m\times n\times T}$ based on each fund’s average investment proportion in each stock over multiple historical years. For each temporal slice $t\in \{1,\dots ,T\}$, we define the holding matrix $M_t\in {\mathbb{R}}^{m\times n}$, where m and n denote the number of funds and stocks, respectively. Each tensor element $M_{ij}^t$ represents the holding weight of fund i in stock j at date t.

To extract fundamental stock features, we perform dimensionality reduction on $\mathcal{M}$ through matrix factorization, decomposing it into low-rank matrices: a fund preference matrix $F\in {\mathbb{R}}^{m\times k}$ and a stock fundamental attribute matrix $Q\in {\mathbb{R}}^{n\times k}$, where $k\ll min(m,n)$ indicates the latent feature dimension. Each stock j is associated with a vector $q_j\in {\mathbb{R}}^k$, while each fund i corresponds to $fi\in {\mathbb{R}}^k$. A fund i’s investment behavior toward stock j can be approximated via the similarity function $\widehat{\gamma }i,j=f_i^{\top }{q}_j$ in the latent space, where the inner product aggregates fund i’s overall preference for stock j across various latent attribute dimensions.

To model prior biases inherent in real-world stock investments, we introduce fund-specific bias ${\mu }_f^{t,i}$, stock-specific bias ${\mu }_q^{t,j}$, and global bias ${\mu }^t$, respectively capturing individual deviations at the fund, stock, and model levels. The latent representations $f_i^t$ and $q_j^t$, which encode fundamental characteristics of funds and stocks, are estimated by minimizing the regularized objective function:

$$\underset{\theta }{min}\sum _{(i,j)}{\left(M_{ij}^t-{\mu }^t-{\mu }_f^{t,i}-{\mu }_q^{t,j}-f_i^{t\top }{q}_j^t\right)}^2+\lambda \parallel \theta \parallel$$

(3)

where $\parallel \theta \parallel =\left(|f_i^t{|}^2+{\left|q_j^t\right|}^2+{\left({\mu }_f^{t,i}\right)}^2+{\left({\mu }_q^{t,j}\right)}^2\right)$, with the regularization term $\lambda \parallel \theta \parallel$ mitigating overfitting.

3.3. Market-Mainline Encoding

Given the dynamic evolution of market preferences for distinct stock attributes across different time periods, market mainlines prove elusive to manual definition. To address this, we propose a dynamically optimized stock-set aggregation algorithm with adaptive refinement. This approach extracts latent market mainlines and their representations from a specific stock-set while establishing dynamic stock–mainline connections, thereby leveraging mainline-related features derived from individual stocks to iteratively optimize mainline representations. The procedure comprises two sequential stages.

3.3.1. Market-Mainline Initialization via High-Return Stocks

Given that top-performing stocks reflect prevailing market preferences, we propose an initialization method that aggregates features from the highest-return stocks to construct representations of upward market mainlines. The algorithm proceeds as follows:

(1)

For each trading day t, the top-n stocks by return are selected to form the set $S_{\mathrm{top} n}^t$, which represents the core stocks driving market gains. We initialize representations of multiple coexisting market mainlines for the current day, assuming n mainlines corresponding to the n stocks. Using each stock i’s embeddings $p_i^t$ and $q_i^t$, initialize the embeddings ${mp}_i^{t,0}$ and ${mq}_i^{t,0}$ for temporal mainline ${MP}_i$ and fundamental mainline ${MQ}_i$, where index i corresponds to return rankings at t.

(2)

Compute cosine similarities between all stocks and market mainlines for both temporal and fundamental features:

$${\alpha }_{xi}^{t,0}=\mathrm{Cosine}(p_i^t,m{p}_x^{t,0})={\displaystyle \frac{p_i^t·m{p}_x^{t,0}}{\left|\right|p_i^t\left|\right|·\left|\right|m{p}_x^{t,0}\left|\right|}}$$

(4)

$${\beta }_{yi}^{t,0}=\mathrm{Cosine}(q_i^t,m{q}_y^{t,0})={\displaystyle \frac{q_i^t·m{q}_y^{t,0}}{\left|\right|q_i^t\left|\right|·\left|\right|m{q}_y^{t,0}\left|\right|}}$$

(5)

where $\alpha {xi}^{t,0}$ and $\beta {yi}^{t,0}$ represent the cosine similarity between stock i and temporal market mainline $M{P}_x$/fundamental market mainline $M{Q}_y$, respectively. The initialization and similarity computation process is illustrated in Figure 3a.

(3)

Stocks are assigned to their most similar market mainlines, excluding the mainline they initially generated. Mainlines unconnected to any stocks are pruned. This process is illustrated in Figure 3b; stocks $p_1^t$, $p_2^t$ and $p_3^t$ are assigned to their most similar market mainline ${mp}_2^{t,0}$, ${mp}_2^{t,0}$ and ${mp}_3^{t,0}$, respectively, while ${mp}_1^{t,0}$ is pruned due to the absence of any connecting stock.

(4)

For stocks whose originally initialized mainline remains after pruning, the connection to their counterpart is retained.

(5)

Using the stock-to-mainline cosine similarities ${\alpha }^{t,0}$ and ${\beta }^{t,0}$ as aggregation weights, we compute updated representations $m{p}_x^{t,1}$ and $m{q}_y^{t,1}$ for temporal market mainline $M{P}_x$ and fundamental market mainline $M{Q}_y$:

$$m{p}_x^{t,1}=\sum _{i\in {\mathcal{M}}_x^t}{\alpha }_{xi}^{t,0}{p}_i^t$$

(6)

$$m{q}_y^{t,1}=\sum _{i\in {\mathcal{C}}_y^t}{\beta }_{yi}^{t,0}{q}_i^t$$

(7)

where ${\mathcal{M}}_x^t$ denotes the set of stocks connected to temporal market mainline $M{P}_x$, and ${\mathcal{C}}_y^t$ represents the stock set associated with fundamental market mainline $M{Q}_y$. The aggregation and representation update process is shown in Figure 3c.

3.3.2. Market Mainline Representation Refinement

Relying solely on the highest-return stock cohort for market mainline construction presents two limitations:

(1)

Stocks highly correlated with market mainlines may be omitted due to insufficient return rankings, resulting in incomplete mainline representations.

(2)

Certain high-return stocks may lack substantive mainline relevance, and their inclusion could introduce noise during mainline formation.

To address these limitations, we refine mainline representations by leveraging implicit associations between all stocks and mainlines. Stocks exhibiting exceptionally high similarity to mainlines despite sub-threshold returns are incorporated as latent missing components. Conversely, stocks with low similarity to initial mainlines, even with high returns, are excluded as non-contributing elements.

To quantify the degree of association between individual stocks and the market mainlines, we compute the cosine similarity ${\gamma }_{xi}^{t,0}$ between each stock’s temporal embedding $p_i^t$ and the initial representation ${mp}_x^{t,1}$ of temporal mainline, and the cosine similarity ${\delta }_{xi}^{t,0}$ between each stock’s fundamental embedding $q_i^t$ and the initial representation ${mq}_y^{t,1}$ of fundamental mainline:

$${\gamma }_{xi}^{t,0}=\mathrm{Cosine}(p_i^t,m{p}_x^{t,1})={\displaystyle \frac{p_i^t·m{p}_x^{t,1}}{\left|\right|p_i^t\left|\right|·\left|\right|m{p}_x^{t,1}\left|\right|}}$$

(8)

$${\delta }_{yi}^{t,0}=\mathrm{Cosine}(q_i^t,m{q}_y^{t,1})={\displaystyle \frac{q_i^t·m{q}_y^{t,1}}{\left|\right|q_i^t\left|\right|·\left|\right|m{q}_y^{t,1}\left|\right|}}$$

(9)

Subsequently, normalize the cosine similarities via softmax to obtain aggregation weights ${\gamma }_{xi}^{t,1}$ and ${\delta }_{yi}^{t,1}$:

$${\gamma }_{xi}^{t,1}={\displaystyle \frac{exp\left({\gamma }_{xi}^{t,0}\right)}{{\sum }_{j\in S_t}exp\left({\gamma }_{xj}^{t,0}\right)}}$$

(10)

$${\delta }_{yi}^{t,1}={\displaystyle \frac{exp\left({\delta }_{yi}^{t,0}\right)}{{\sum }_{j\in S_t}exp\left({\delta }_{yj}^{t,0}\right)}}$$

(11)

The market mainline representations are then refined through weighted aggregation across all stocks:

$$m{p}_x^{t,2}=\mathrm{LeakyReLU}\left(W_u\left(\sum _{i\in {\mathcal{S}}_t}{\gamma }_{xi}^{t,1}{p}_i^t\right)+b_u\right)$$

(12)

$$m{q}_y^{t,2}=\mathrm{LeakyReLU}\left(W_u\left(\sum _{i\in {\mathcal{S}}_t}{\delta }_{yi}^{t,1}{q}_i^t\right)+b_u\right)$$

(13)

This dynamic refinement, as shown in Figure 4a, enhances the robustness and interpretability of market mainline representations, where $W_u$ and $b_u$ denote learnable parameters, LeakyReLU serves as the activation function, and the aggregation weights ${\gamma }_{xi}^{t,1}$ and ${\delta }_{yi}^{t,1}$ resolve the aforementioned two limitations.

3.4. Heterogeneous Dynamic Effects of Market Mainlines on Stocks

Building upon the established modeling of stock-specific and market-mainline features, this section addresses the core question: How do dynamically evolving market mainlines exert differential effects across individual stocks. Fundamental characteristics exhibit relative stability, reflecting intrinsic value and corporate operational essence, whereas temporal features capture dynamic market sentiments and short-term expectations. To resolve the inherent limitation of static stock attributes in predicting trends within transient market dynamics, we employ distinct computational strategies to model the effects of temporal and fundamental mainlines on individual stocks, thereby enabling dynamic resolution of their heterogeneous influences.

3.4.1. Temporal Mainline Effects on Stocks

To explore the differential impact of various market mainlines on individual stocks, and considering that different mainlines exhibit varying levels of influence (where some mainlines may be more significant while others are less impactful), the importance of each market temporal mainline for each stock is first learned. The representations of market temporal mainlines are then aggregated according to these learned importance weights to capture the influence of each market mainline on individual stocks.

Specifically, the cosine similarity ${\epsilon }_{xi}^{t,0}$ between the temporal embedding of stock $p_i^t$ and the representation of market temporal mainline $m{p}_x^{t,2}$ is computed. Subsequently, similarity values between all temporal mainlines in set $M{P}_{t,2}$ and stock i are normalized via softmax, thus deriving the influence weight ${\epsilon }_{xi}^{t,1}$ from temporal market mainline representation $m{p}_x^{t,2}$ to stock i:

$${\epsilon }_{xi}^{t,0}=\mathrm{Cosine}(p_i^t,m{p}_x^{t,2})={\displaystyle \frac{p_i^t·m{p}_x^{t,2}}{\left|\right|p_i^t\left|\right|·\left|\right|m{p}_x^{t,2}\left|\right|}}$$

(14)

$${\epsilon }_{xi}^{t,1}={\displaystyle \frac{exp\left({\epsilon }_{xi}^{t,0}\right)}{{\sum }_{x\in M{P}_{t,2}}exp\left({\epsilon }_{xi}^{t,0}\right)}}$$

(15)

Finally, the aggregated information ${\widehat{sp}}_i^{t,0}$, representing the combined influence of all market temporal mainlines on stock i at date t, is obtained by performing a weighted aggregation of the influence information from each market mainline. This aggregated representation is then fed into a fully connected layer with a LeakyReLU activation function:

$${\widehat{sp}}_i^{t,0}=\mathrm{LeakyReLU}\left(W_s\left(\sum _{x\in \mathcal{M}{P}_{t,2}}{\epsilon }_{xi}^{t,1}{p}_i^t\right)+b_s\right)$$

(16)

The process of extracting and aggregating the influence of various market mainlines on an individual stock is visually summarized in Figure 4b.

3.4.2. Fundamental Mainline Effects on Stocks

Since fundamental attributes of individual stocks remain static over given periods, these invariant properties cannot explicitly predict stock trends in dynamic markets. To address this limitation, we dynamically model the aggregate characteristics of fundamental market mainlines and compute their correlations with stock-specific fundamental representations, thereby capturing the influence of market-wide fundamental mainlines on individual equities.

Specifically, given the refined fundamental mainline set $M{Q}_{t,2}$, average pooling is applied to derive the aggregated market fundamental mainline $m{q}^{t,3}$. Subsequently, the cosine similarity ${\eta }_i^{t,0}$ between each stock’s fundamental embedding $q_i^t$ and the aggregated market fundamental mainline feature is computed. These similarity values are normalized across all stocks in set $S_t$ via softmax to quantify the dynamic influence ${\widehat{sq}}_i^t$ of fundamental market mainlines on stock i:

$$m{q}^{t,3}=\mathrm{AvgPool}\left({\left\{m{q}_y^{t,2}\right\}}_{y\in M{Q}_{t,2}}\right)$$

(17)

$${\eta }_i^{t,0}=\mathrm{Cosine}(q_i^t,m{q}^{t,3})={\displaystyle \frac{q_i^t·m{q}^{t,3}}{\left|\right|q_i^t\left|\right|·\left|\right|m{q}^{t,3}\left|\right|}}$$

(18)

$${\widehat{sq}}_i^{t,0}={\displaystyle \frac{exp\left({\eta }_i^{t,0}\right)}{{\sum }_{j\in S_t}exp\left({\eta }_j^{t,0}\right)}}$$

(19)

3.5. Stock-Specific Individual Information

Stock price fluctuations are governed by both market-wide patterns and individual dynamics. To precisely capture distinctive characteristics of individual equities, we extract temporal information independent of market mainline influences—referred to as stock-specific individual information—which characterizes unique evolutionary patterns. These individual price dynamics are captured through residual computation between stock i’s temporal feature $p_i^t$ and the mainline-influenced representation ${\widehat{sp}}_i^{t,1}$, with the residual processed through a fully connected layer featuring LeakyReLU activation:

$${\widehat{sp}}_i^{t,1}=\mathrm{LeakyReLU}\left(W_b{\widehat{sp}}_i^{t,0}+b_b\right)$$

(20)

$${\widehat{sd}}_i^{t,0}=\mathrm{LeakyReLU}\left(W_i\left(p_i^t-{\widehat{sp}}_i^{t,1}\right)+b_i\right)$$

(21)

3.6. Prediction Module

To preserve information integrity and enhance feature expressiveness, the temporal mainline influence ${\widehat{sp}}_i^{t,0}$ and stock-specific dynamics ${\widehat{sd}}_i^{t,0}$ are element-wise summed at date t. This sum is concatenated with the fundamental mainline influence ${\widehat{sq}}_i^{t,0}$, forming the predictive input ${\widehat{s}}_i^t$. Concatenation preserves the complete information structures of both temporal and fundamental mainline influences. The fused representation is then processed through a fully connected layer to generate stock i’s trend prediction $r_i^t$:

$$r_i^t=W_r{s}_i^t+b_r=W_r\left([{\widehat{sp}}_i^{t,0}+{\widehat{sd}}_i^{t,0}],{\widehat{sq}}_i^{t,0}\right)+b_r$$

(22)

3.7. Training Objective

The proposed model is trained using the Adam optimization algorithm, minimizing the Mean Squared Error loss function to optimize parameters. The loss function is defined as

$$\mathcal{L}=\sum _{t\in \mathcal{T}}\mathrm{MSE}(r^t,d^t)=\sum _{t\in \mathcal{T}}\sum _{i\in {\mathcal{S}}^t}{\displaystyle \frac{{\left(r_i^t-d_i^t\right)}^2}{|{\mathcal{S}}^t|}}$$

(23)

where $\mathcal{T}$ denotes the set of trading dates in the training period, ${\mathcal{S}}^t$ represents the stock universe at date t, $r_i^t$ is the predicted trend, and $d_i^t$ is the actual observed trend for stock i at time t. Minimization of $\mathcal{L}$ enables the model to extract predictive capability for stock trends from historical patterns.

4. Experiments

4.1. Experimental Setup

4.1.1. Datasets

Stock Sets

This study employs Chinese A-share market data from 2013 to 2024, which includes daily price and trading-volume time series for individual stocks. The raw data are sourced from the open-source quantitative finance database Qlib (https://github.com/microsoft/qlib (accessed on 1 February 2026)) [44]. To ensure data cleanliness and consistency, we remove stocks that were labeled as ST (Special Treatment) or *ST (delisting-risk warning) during the sample period.

To systematically assess the performance of the proposed Market-mainline-Driven Multi-feature Fusion Model (MDMF), we employ three datasets that differ distinctly in market size, liquidity, and heterogeneity: the constituents of the CSI 300 index, the constituents of the CSI 500 index, and a full-market stock set (with ST/*ST stocks excluded). Specifically, the CSI 300 constituents represent the market’s largest and most liquid core assets, whose price movements are primarily driven by macroeconomic and systematic factors. This dataset serves to test the model’s ability to identify market mainlines in a high-signal environment. The CSI 500 constituents comprise more diverse mid-cap growth stocks, which are influenced by both systematic factors and stronger stock-specific noise. This dataset challenges the model to distinguish common market signals from idiosyncratic disturbances. Finally, the full-market universe provides a highly heterogeneous environment, testing the model’s scalability, robustness, and generalizability under complex, real-world conditions. The key attributes of these three datasets are summarized in

Table 1. Key attributes of the experimental datasets.

Attribute	CSI 300	CSI 500	Full-Market (Excl. ST)
Number of Stocks	300	500	2146 to 3959
Industry Coverage	Broad, covering major sectors	Broad, covering major sectors	Broad, covering all market sectors
Market Cap Range	Large-cap	Mid- and Small-cap	Full range
Market	China A-Share	China A-Share	China A-Share
Market Cap Coverage	Covers about 70% of total A-share market cap	Covers about 15% of total A-share market cap	Covers the vast majority of non-ST stocks
Start Time	1 September 2013	1 September 2013	1 September 2013
End Time	31 August 2024	31 August 2024	31 December 2020

for clear comparison.

Two distinct time frames are employed to assess the model’s robustness under varying market conditions, taking into account the impact of market bull–bear cycles and “black swan” events such as the COVID-19 pandemic. The first period covers September 2013 to December 2020, where the first 75% of samples are used for training, the next 25% for validation, and the remaining final segment for testing. The second period extends the data to August 2024, with the same proportional split applied. It should be noted that the full-market dataset is not extended beyond 2020. This is due to the substantial data noise and missing values introduced by the rapid market expansion between 2021 and 2024, which included over 1,400 new listings and would obscure a clear evaluation. Detailed descriptions of all datasets are provided in

Table 2. Basic information about the experimental datasets.

Dataset	Data Type	Start Time	End Time	Days	Data Volume
	Train	1 September 2013	31 August 2017	976	268,084
CSI 300 (2013–2020)	Valid	1 September 2017	31 December 2018	334	93,208
	Test	1 January 2019	31 December 2020	481	125,539
	Train	1 September 2013	31 August 2019	1463	419,950
CSI 300 (2013–2024)	Valid	1 September 2019	31 August 2021	486	145,524
	Test	1 September 2021	31 August 2024	728	217,979
	Train	1 September 2013	31 August 2017	976	433,432
CSI 500 (2013–2020)	Valid	1 September 2017	31 December 2018	334	152,705
	Test	1 January 2019	31 December 2020	481	208,973
	Train	1 September 2013	31 August 2019	1463	688,224
CSI 500 (2013–2024)	Valid	1 September 2019	31 August 2021	486	242,473
	Test	1 September 2021	31 August 2024	728	362,466
	Train	1 September 2013	31 August 2017	976	2,344,660
Full-market (2013–2020)	Valid	1 September 2017	31 December 2018	334	1,067,545
	Test	1 January 2019	31 December 2020	481	1,531,085

.

Stock Features

Alpha360 stock features from the open-source quantitative investment platform Qlib (https://github.com/microsoft/qlib (accessed on 1 February 2026)) [44] were used. Daily features include six data points: opening price, closing price, high price, low price, volume-weighted average price (VWAP), and trading volume. For each stock at date t, Alpha360 constructs a 360-dimensional feature vector using historical data from the past 60 days.

Mutual Fund Semi-Annual Portfolio Reports

Semi-annual investment portfolio reports for Chinese mutual funds from Q3 2013 to Q4 2020 were collected from East Money (https://www.eastmoney.com/ (accessed on 1 February 2026)). These reports provide biannual investment amounts in each stock.

4.1.2. Parameter Settings

All models in this study were implemented and trained using the PyTorch 2.1.2 framework. To ensure full reproducibility, the random seeds for both numpy and torch were fixed at 0 at the beginning of each experiment. Each training and testing batch was set to include stock features from the same trading day. Consequently, the batch size for both the Market-Mainstream-Driven Multi-feature Fusion Model and baseline models equals the total number of stocks on that day. The temporal feature encoder employs a 2-layer GRU with a hidden dimension of 64. The learning rate was fixed at 0.002. Each model was trained for a maximum of 200 epochs with an early stopping strategy, which halted training if the validation set’s IC score did not improve for 30 consecutive epochs.

The dynamic stock aggregation for market-mainline initialization was configured as follows: on each trading day, stocks ranked in the top 2% by return within the daily stock pool were selected to construct the initial upward market-mainline representation. This set of top-performing stocks reflects prevailing market preferences and serves as the input for the subsequent aggregation algorithm detailed in Section 3.3.1.

To construct stable fundamental stock representations, semi-annually published holding reports of Chinese publicly offered funds were employed. Regulatory requirements mandate that semi-annual reports be released before September 1st each year, and annual reports before April 1st of the following year. At each prediction time point, we aggregated investment behaviors from the three most recent consecutive reports to form the fund-stock holding matrix $M_t$. The stock representations were updated in April and September, aligning with the report release cycle to prevent the use of future information. For instance, for the period from 2 September 2013 to 31 March 2014, we utilized the semi-annual report released before September 2012, the annual report before April 2013, and the semi-annual report before September 2013. For learning fundamental stock features, a grid search strategy systematically optimized hyperparameters of the matrix factorization model to enhance precision and generalization capability in extracting intrinsic stock attributes. Specifically, the latent vector length was set to 64, training epochs were determined from $\{10,15,20,50,100\}$, learning rates were adjusted within $\{0.002,0.005,0.01\}$ to find suitable parameter update step sizes, regularization strength $\lambda$ was selected from $\{0.1,1,10\}$ to effectively control model complexity and prevent overfitting.

4.1.3. Compared Methods

Our Market-Mainstream-Driven Multi-feature Fusion Model was benchmarked against the following stock trend prediction methods:

• GRU [45]: GRU is a variant of RNN with gating mechanisms that effectively addresses long-term dependency issues when learning sequential features from time-series data, mitigating gradient vanishing and explosion problems. It has been widely applied in stock prediction.
• IMSR [11]: IMSR integrates dynamic stock representations with their correlations to market states, using MLPs to predict next-period stock rankings. It dynamically captures market preferences based on market representations to enhance stock prediction.
• GATs_ts [46]: A dynamic graph attention network that feeds GRU-encoded stock embeddings into GATs. It aggregates these embeddings on stock graphs using graph attention networks to capture dynamic dependencies between stocks.
• FinGAT [47]: A graph neural network-based model that constructs fully connected graphs to learn intra-industry relationships, employs graph pooling mechanisms to generate industry embeddings, and combines graph attention networks to learn inter-industry relationships. It models hierarchical relationships between stocks and industries to learn mutual influences.
• HIST [48]: A graph-based stock trend prediction framework that fully utilizes shared information in graphs constructed from predefined and hidden concepts by mining concept-oriented shared information.
• MTMD [49]: An extension of the HIST model incorporating learnable embeddings, external attention mechanisms, and memory modules. It adaptively fuses heterogeneous multi-scale information through graph networks to improve prediction accuracy.
• ESTIMATE [50]: Combines temporal generative filters with hypergraph convolution. It uses temporal generative filters on LSTM networks to learn stock-specific patterns via memory-based mechanisms, while employing hypergraph attention to capture non-paired correlations. Wavelet bases simplify message passing to focus on local convolution, addressing efficient convolution of high-order dynamics.
• STHAN-SR [51]: Utilizes Hawkes process-based temporal attention LSTM to model stock relationships. It constructs hypergraphs from inter-stock relationships and employs designed spatial hypergraph convolution modules with attention mechanisms to capture spatio-temporal dependencies, generating stock rankings.
• StockMixer [52]: An MLP-based stock price prediction architecture that processes through three stages: indicator mixing, temporal mixing, and stock mixing. It efficiently captures complex relationships in stock data across indicators, time, and inter-stock associations.

4.1.4. Metrics

To enhance the reliability of the evaluation, we conduct ten independent training and prediction runs for each method and use the average result as the final assessment. The primary objectives are to evaluate the performance of the baseline model and our proposed model, and to effectively identify stocks with the highest predictive value or return potential. For this purpose, we employ the following six key metrics for a multidimensional evaluation: the Information Coefficient (IC) [53], the Rank Information Coefficient (Rank IC) [54], the Information Ratio (ICIR), the Rank Information Ratio (Rank ICIR), the Sharpe Ratio, and Precision@N. The definitions and calculations of these metrics are as follows.

• Information Coefficient (IC): This metric measures the linear correlation between the model’s predicted values and the actual directional labels of stock returns. It is calculated as the Pearson correlation coefficient between these two sets of values across all trading days and for each individual stock.
• Rank Information Coefficient (RIC): This metric assesses the consistency between the ranking order of model predictions and the ranking order of actual labels. It is calculated as the Spearman’s rank correlation coefficient across all trading days and stocks, providing a more robust measure of monotonic relationship.
• Information Ratio (ICIR): This ratio evaluates the stability and significance of the Information Coefficient. It is calculated as the ratio of the mean IC to its standard deviation, typically annualized. The formula is $ICIR={\displaystyle \frac{\mathrm{Mean}\left(IC\right)}{\mathrm{Std}\left(IC\right)}}$. A higher ICIR indicates more consistent and reliable predictive ability of the model.
• Rank Information Ratio (RICIR): This metric is analogous to the ICIR but is calculated based on the Rank IC to assess the stability and significance of ranking prediction capability. Its annualized formula is $RankICIR={\displaystyle \frac{\mathrm{Mean}\left(RankIC\right)}{\mathrm{Std}\left(RankIC\right)}}$.
• Sharpe Ratio (SR): This metric measures the risk-adjusted return of a daily portfolio constructed from the model’s top-2% predicted stocks (equally weighted). Assuming a zero risk-free rate and 252 trading days per year, the annualized Sharpe Ratio is computed as $\widehat{\mathrm{SR}}={\displaystyle \frac{\sqrt{252}·{\widehat{\mu }}_d}{{\widehat{\sigma }}_d}}$, where ${\widehat{\mu }}_d$ and ${\widehat{\sigma }}_d$ are the sample mean and standard deviation of the daily portfolio returns.
• Precision@N: This metric evaluates the model’s accuracy in identifying stocks with the highest potential for price increase. Specifically, on each trading day, all stocks are ranked in descending order based on the model’s prediction scores. The top N stocks are selected, and the proportion of these with true positive labels (i.e., actual price increase) is calculated. We report the daily average Precision@N for $N=3,5,10,$ and 30.

4.2. Main Results

Table 3. Results on CSI 300 (2013–2020) dataset.

Methods	IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@N (↑)
						3	5	10	30
GRU	0.085	0.082	0.714	0.689	4.64	56.93	56.70	56.44	55.94
HIST	0.105	0.103	0.763	0.741	5.54	58.17	57.96	57.62	57.32
IMSR	0.088	0.084	0.733	0.704	5.04	56.27	56.09	55.85	55.49
GATs	0.086	0.082	0.766	0.723	5.27	56.92	56.56	55.83	55.69
FinGAT	0.092	0.090	0.701	0.685	5.61	57.11	56.80	56.71	55.99
MTMD	0.103	0.101	0.732	0.705	5.27	57.46	57.23	57.88	56.86
StockMixer	0.087	0.084	0.735	0.710	5.14	56.76	56.53	55.91	55.64
STHAN-SR	0.095	0.090	0.751	0.712	5.06	58.01	57.84	57.33	56.97
ESTIMATE	0.097	0.091	0.747	0.709	4.95	60.12	59.49	58.04	54.77
MDMF	0.108	0.106	0.801	0.777	5.96	59.11	58.59	58.18	57.45

The upward arrow (↑) indicates that higher values are better for that metric. Bold values indicate the highest scores for each metric.

,

Table 4. Results on CSI 500 (2013–2020) dataset.

Methods	IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@N (↑)
						3	5	10	30
GRU	0.105	0.096	1.004	0.921	5.61	58.05	57.03	56.03	55.74
HIST	0.121	0.115	1.031	0.962	6.01	58.17	57.67	57.02	57.07
IMSR	0.111	0.100	1.046	0.947	6.24	58.40	57.62	57.14	56.03
GATs	0.104	0.097	0.994	0.915	5.66	57.42	56.23	55.07	55.04
FinGAT	0.113	0.104	1.031	0.948	5.87	56.51	55.81	55.60	55.48
MTMD	0.120	0.115	1.004	0.942	5.79	57.96	55.90	55.54	56.11
StockMixer	0.106	0.097	1.015	0.924	5.32	57.05	56.28	55.70	55.49
STHAN-SR	0.116	0.109	0.986	0.915	5.68	58.17	58.00	57.04	55.86
ESTIMATE	0.114	0.108	0.930	0.876	5.70	57.98	56.78	55.55	55.26
MDMF	0.124	0.117	1.050	0.971	6.38	58.87	58.36	57.12	56.73

and

Table 5. Results on full-market (2013–2020) stock set.

Methods	IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@N (↑)
						3	5	10	30
GRU	0.154	0.134	1.626	1.350	18.39	90.76	86.37	78.02	64.58
HIST	0.162	0.150	1.699	1.485	18.93	91.88	87.56	79.62	66.94
IMSR	0.153	0.135	1.691	1.437	19.26	89.92	85.26	77.99	66.39
GATs	0.149	0.131	1.683	1.399	17.65	90.74	86.11	77.31	64.45
FinGAT	0.158	0.135	1.705	1.392	18.28	89.92	85.26	77.99	66.39
MTMD	0.161	0.145	1.729	1.435	18.76	91.87	86.56	79.27	66.85
StockMixer	0.155	0.139	1.699	1.394	17.49	87.96	84.74	78.10	64.81
STHAN-SR	0.158	0.141	1.777	1.463	18.73	91.73	86.62	78.95	65.62
ESTIMATE	0.159	0.141	1.768	1.456	18.80	91.95	87.64	79.63	66.37
MDMF	0.168	0.151	1.737	1.471	19.68	92.06	87.88	79.91	67.34

present the empirical performance of the proposed Market-mainline-Driven Multi-feature Fusion Model (MDMF), evaluated on the CSI 300, CSI 500, and full-market datasets. The model was trained on data from 2013 to 2020 and tested during the bull market period of 2019–2020, with comparisons made against a series of benchmark models. The results indicate that our MDMF model demonstrates robust and consistently superior performance across stock universes with varying characteristics.

On the less volatile, large-cap CSI 300 as well as the more volatile, mid/small-cap CSI 500, it achieves the highest scores in IC, Rank IC, IR, Rank IR, and Sharpe Ratio, while maintaining competitive Precision@N. These results collectively underscore the model’s robust capacity to capture fundamental pricing logic. More importantly, MDMF effectively translates this understanding into both superior ranking accuracy and exceptional risk-adjusted returns. While MDMF does not attain the highest ICIR or Rank ICIR on the more diverse and noisy full-market stock set, it secures statistically significant leads in IC, Rank IC and most notably the Sharpe Ratio. This indicates that our model successfully identifies high-potential stocks within a broad market context, delivering superior risk-adjusted performance even amidst greater market noise.

Although MDMF does not achieve the highest Information Ratio (IR) or Rank ICIR on the more diverse and noisier full-market stock universe, it attains statistically significant leads in the Information Coefficient (IC), Rank IC, and, most notably, the Sharpe Ratio. To further validate the stability of the model’s performance, we computed 95% confidence intervals for these key metrics across the three datasets. Specifically, the intervals for IC, IR, and the Sharpe Ratio on the CSI 300 dataset are (0.1055, 0.1093), (0.1036, 0.1073), (5.724, 6.195); on the CSI 500 dataset, they are (0.1216, 0.1259), (0.1142, 0.1192), (5.966, 6.811); and on the full-market dataset, they are (0.1655, 0.1698), (0.1487, 0.1532), (19.426, 19.939). These quantified uncertainty assessments demonstrate that our model can robustly identify high-potential stocks in broader market environments, delivering superior and statistically reliable risk-adjusted performance even amid heightened market noise.

To intuitively assess the return and risk characteristics of the portfolio constructed by the model, we analyze the distribution of daily returns based on its predictions (by selecting the top 2% of stocks ranked by prediction each day to form an equally weighted portfolio). As shown in Figure 5, during the 2013–2020 test period, the daily return distributions for the CSI 300 and CSI 500 portfolios exhibit an approximately normal shape with significantly positive means, consistent with their high Sharpe Ratios. The full-market portfolio shows an even higher positive mean and a narrower distribution range, explaining its outstanding Sharpe Ratio performance.

However, during the 2021–2024 period (see

Table 6. Results on CSI 300 (2013–2024) dataset.

Methods	IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@N (↑)
						3	5	10	30
GRU	0.064	0.062	0.523	0.501	1.85	52.07	51.96	52.36	52.83
HIST	0.067	0.064	0.457	0.434	2.77	54.61	54.13	53.93	54.11
IMSR	0.066	0.064	0.447	0.426	2.72	55.70	55.23	54.18	54.27
GATs	0.053	0.050	0.427	0.394	1.48	52.21	52.28	52.40	52.49
FinGAT	0.058	0.056	0.439	0.417	1.98	53.59	52.92	53.27	53.41
MTMD	0.067	0.065	0.454	0.431	2.78	54.63	54.91	54.78	53.99
StockMixer	0.059	0.057	0.438	0.409	1.52	52.08	51.49	52.32	52.64
STHAN-SR	0.068	0.066	0.486	0.465	2.39	53.65	53.11	52.62	52.89
ESTIMATE	0.069	0.067	0.536	0.518	2.10	53.18	53.25	53.27	53.91
MDMF	0.070	0.067	0.472	0.449	2.93	54.11	54.13	53.94	53.80

and

Table 7. Results on CSI 500 (2013–2024) dataset.

Methods	IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@N (↑)
						3	5	10	30
GRU	0.068	0.064	0.558	0.531	2.17	52.59	53.37	53.28	52.64
HIST	0.072	0.069	0.675	0.618	2.31	54.26	53.63	53.51	53.72
IMSR	0.071	0.067	0.538	0.509	1.98	54.74	53.06	53.31	53.83
GATs	0.059	0.056	0.516	0.487	1.69	52.56	52.39	52.36	52.26
FinGAT	0.063	0.061	0.527	0.499	2.20	53.67	52.78	52.56	52.31
MTMD	0.070	0.064	0.682	0.627	2.20	54.04	53.43	53.08	53.32
StockMixer	0.062	0.058	0.439	0.404	1.43	53.46	52.51	51.99	52.70
STHAN-SR	0.074	0.071	0.568	0.543	2.66	53.18	53.35	53.43	53.90
ESTIMATE	0.073	0.069	0.571	0.535	2.39	53.82	52.90	53.59	54.03
MDMF	0.071	0.068	0.547	0.518	2.79	54.24	54.11	53.96	54.06

), the model’s performance metrics generally decline, particularly for the CSI 500 index. This phase coincides with market downturns and consolidation, and the corresponding daily return distributions (Figure 6) show notable changes: the mean shifts toward zero, the distribution widens, and the proportion of negative returns increases. These results indicate that the model’s performance depends on the market environment: during upward trends, clear market mainlines and capital consensus provide effective signals for the model. In contrast, during downturns or periods of panic, market mainlines become blurred and shift rapidly, making it more difficult for the model to capture such signals. This effect is particularly pronounced for mid- and small-cap stocks, which are more susceptible to heterogeneous noise, thereby impacting the model’s overall performance.

Despite this dependence on market regimes, MDMF’s overall advantage across multiple market segments and evaluation dimensions remains solid compared to existing benchmarks. Hypergraph-based models (ESTIMATE, STHAN-SR) implicitly reflect partial market attributes through industry information aggregation; their essence remains an expression constrained by predefined structures, primarily focusing on complex inter-stock relationship modeling. While hyperedge-based weighted averaging of industry features captures group commonality, it inadequately characterizes core market-driving logic—such as policy orientation and capital preferences—which constitute dominant dynamic factors. Compared to hypergraph architectures, market mainline information more precisely captures deep logical correlations between holistic market attributes and individual stocks, strengthening the model’s capability to comprehend overall market movements and consequently delivering superior predictive performance.

Furthermore, although IMSR captures dynamic interactions between stocks and market mainlines, it relies solely on static market-mainline feature construction and explores stock-market relationships through a single dimension. In contrast, the proposed MDMF model dynamically quantifies cross-dimensional correlations between stocks and market mainlines, capturing more complex interdependencies. Consequently, this approach outperforms existing LSTM-based IMSR methods.

4.3. Ablation Study

To systematically evaluate the contribution of each key component in our proposed MDMF framework, we conducted a series of ablation studies aimed at addressing three core aspects: the effectiveness of the market-mainline initialization and refinement strategy, the respective necessity of temporal and fundamental information in capturing heterogeneous market effects, and the importance of stock-specific individual information. The results across the CSI 300 (2013–2020) and CSI 500 (2013–2020) are summarized in

Table 8. Ablation study results on the contribution of key components in the MDMF framework.

Temporal	Fundamental	Individual	Metrics
			IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@30 (↑)
CSI 300 (2013–2020)
✓	-	-	0.091	0.088	0.731	0.711	4.92	56.52
-	✓	-	0.089	0.086	0.741	0.712	5.04	55.65
✓	✓	-	0.092	0.089	0.749	0.723	4.35	56.67
✓	✓	✓	0.108	0.106	0.801	0.777	5.96	57.45
CSI 500 (2013–2020)
✓	-	-	0.108	0.100	0.984	0.908	5.40	55.85
-	✓	-	0.111	0.101	1.046	0.947	6.24	56.03
✓	✓	-	0.106	0.098	0.951	0.880	4.92	55.60
✓	✓	✓	0.124	0.117	1.050	0.971	6.38	56.73

The ✓ and - indicate having or not having the component in the variants, respectively.

and

Table 9. Performance comparison between initialization-only and refined models for temporal and fundamental market mainlines.

Temporal		Fundamental		Metrics
Initialization	Refinement	Initialization	Refinement	IC (↑)	RIC (↑)	ICIR (↑)	RICIR (↑)	SR (↑)	Precision@30 (↑)
CSI 300 (2013–2020)
✓	-	-	-	0.087	0.085	0.704	0.683	4.69	56.37
✓	✓	-	-	0.091	0.088	0.731	0.711	4.92	56.52
-	-	✓	-	0.089	0.086	0.744	0.714	4.88	55.70
-	-	✓	✓	0.089	0.086	0.741	0.712	5.04	55.65
✓	-	✓	-	0.084	0.083	0.636	0.628	4.46	55.73
✓	✓	✓	✓	0.108	0.106	0.801	0.777	5.96	57.45
CSI 500 (2013–2020)
✓	-	-	-	0.106	0.099	0.980	0.910	5.09	55.48
✓	✓	-	-	0.108	0.100	0.984	0.908	5.40	55.85
-	-	✓	-	0.111	0.101	1.045	0.946	6.04	55.80
-	-	✓	✓	0.111	0.101	1.046	0.947	6.24	56.03
✓	-	✓	-	0.108	0.100	0.909	0.835	5.99	55.48
✓	✓	✓	✓	0.124	0.117	1.050	0.971	6.38	56.73

, and provide comprehensive insights into the role of each component.

(1)

As evidenced by the performance differences in Table 8, both the temporal and fundamental market-mainline information provide distinct and complementary signals for stock trend prediction; removing either component leads to a noticeable performance decline.

(2)

The integration of stock-specific individual information proves indispensable. Its ablation leads to the most significant performance drop across all evaluation metrics. This result underscores that accurate forecasting requires modeling both collective market dynamics and the unique attributes of individual stocks.

(3)

As shown in Table 9, the refined model based on the temporal market mainline significantly outperforms the initialization-only version. This improvement is consistent across all evaluation metrics. It demonstrates that the optimization module plays a critical role in enhancing feature completeness and suppressing noise. For the fundamental market mainline, refinement yields a more nuanced outcome. Although improvements in IC and Rank IC are limited, the SR increases markedly. This indicates that the optimization process effectively improves the risk-adjusted return of the portfolio and helps to build more robust predictive signals.

In conclusion, the ablation studies validate that each component of the MDMF framework—effective market-mainline construction, multi-perspective (temporal and fundamental) effect modeling, and stock-specific modeling—plays a distinct and vital role in achieving superior stock trend forecasting performance.

4.4. Investment Simulation

To further validate the effectiveness of the proposed model, a daily trading-based strategy is implemented for backtesting on both the CSI 300 index and full-market stock datasets. Specifically, a fixed initial capital is allocated at each backtest inception. The investor equally weights the top-K predicted stocks daily, holds selected positions for two trading days, and liquidates stocks exiting the top-K prediction range. Cumulative profits (excluding transaction costs) are reinvested in subsequent trading sessions. Additionally, the average market return is calculated as a baseline reflecting overall market trends, determined by equally weighting all stocks in both datasets. Figure 7 displays cumulative profit curves for all methods at K = 20. The proposed model achieves superior profitability versus all comparative approaches in every backtest. Our model generates cumulative returns exceeding 150% from 2019 to 2020—outperforming the CSI 300 index by 110 percentage points.

5. Conclusions

In this paper, we proposed the Market-mainline-Driven Multi-feature Fusion Model (MDMF), a novel framework that dynamically identifies multidimensional market mainline characteristics and characterizes their differentiated impact on stocks. By integrating a dual-channel encoding mechanism, a dynamic stock set aggregation algorithm, and a differential impact module, the MDMF model effectively leverages both temporal and fundamental perspectives to enhance prediction accuracy. Comprehensive empirical evaluation on real-world market datasets demonstrates that the model outperforms existing benchmark methods in predictive performance and achieves higher cumulative returns in simulated trading environments. However, the study also acknowledges certain limitations of the model: during market downturns or panic phases, when market mainlines tend to become ambiguous, its predictive effectiveness is significantly weaker compared to market uptrends. This phenomenon suggests that the formation and persistence of market mainlines are highly dependent on stable market sentiment and consistent capital consensus, providing important insights into the boundaries of the model’s applicability. Overall, this work not only offers an effective analytical tool for stock prediction but also provides new theoretical perspectives for exploring the formation mechanisms of market mainlines and their impact.

Future research will pursue several directions to enhance the model’s practicality, robustness, and theoretical depth. First, we will systematically investigate the structural sources and potential risks of the model’s returns. This will involve explicitly incorporating key features such as liquidity, market capitalization, dividend policies, and firm-specific events. Coupled with sensitivity analysis and group-wise testing, this approach aims to distinguish returns driven by market mainlines from those attributable to intrinsic asset characteristics or short-term noise, thereby improving the interpretability and robustness of our findings. Second, we plan to extend the research to other financial markets. Testing the model in mature markets (e.g., the U.S. and Europe) as well as in futures and foreign exchange markets will help evaluate its cross-market adaptability. Additionally, we will seek to integrate broader data sources, including macroeconomic variables, news sentiment data, and alternative datasets, to more comprehensively capture the external factors that influence market mainlines.