Authentic SEC Data and Regime-Aware Ensemble Learning for Corporate Cash Flow Forecasting

Abstract

Financial forecasting research often prioritizes methodological sophistication over the authenticity of underlying training data. This study quantifies the “estimation–reality divide” by comparing models trained on estimated quarterly data versus genuine, re-stated SEC-reported cash flows. Using 244 firm-quarter observations from five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta; 2011–2024), this case study shows that, within this specific set of firms, models trained on estimated data exhibit a large optimistic bias. For a state-of-the-art ensemble, this bias appears as a 43% lower error rate (4.5% vs. 7.9%) compared to the same model trained on authentic data. To address this, we introduce a forecasting framework that combines (i) a Hidden Markov Model for detecting economic regimes, (ii) models tailored to each regime (XGBoost and LSTM with attention), and (iii) a dynamic ensemble that adapts to recent performance. In realistic out-of-sample tests, our framework achieves a 7.9% error rate on authentic data, significantly outperforming standard benchmarks. We also show that a meta-learning approach reduces the data needed for a new firm by about 35% while improving accuracy by 24%. In plain terms, using real SEC data leads to more honest and useful forecasts than relying on estimated data. All claims are strictly limited to the five large-cap U.S. technology firms analyzed (Microsoft, Apple, Amazon, Alphabet, Meta). No claims of generalizability to other sectors, firm sizes, or markets are made or implied. Validation on broader samples is required before extending these findings.

1. Introduction

1.1. Motivation: The Data Authenticity Problem in Financial Forecasting

In the field of financial forecasting, remarkable methodological progress has been made, with ensemble methods and deep learning architectures achieving considerable success in forecasting competitions (Makridakis et al., 2020). However, this progress often rests on an under-scrutinized foundation: the quality and authenticity of the training data (Hyndman & Athanasopoulos, 2021). A persistent methodological constraint in corporate finance is the scarcity of high-frequency, genuine financial data. Consequently, a significant body of academic research and practical application relies on estimated quarterly data, often derived by disaggregating annual totals (Dechow et al., 1998; Kim & Kross, 2005).

This practice creates what Huang et al. (2018) call the “estimation–reality divide”. The interpolation and proportional allocation techniques used to create an estimated quarterly series introduce artificial smoothness, allowing for a systematic understating of the true volatility and noise inherent in actual corporate cash flows. For technology firms, characterized by rapid innovation cycles and volatile revenue streams, this reliance on artificially smoothed data is particularly problematic for critical tasks like liquidity management and investment planning (Kumbure et al., 2022). Beyond forecasting accuracy, the concept of data authenticity has gained prominence in recent discussions on data governance and ESG (Environmental, Social, and Governance) reporting, where the integrity of underlying financial data is increasingly viewed as a cornerstone of reliable corporate disclosures (European Commission, 2019; U.S. Department of the Treasury, 2024). The core premise of this study is that methodological advancements cannot compensate for compromised data quality. We argue that progress in financial forecasting requires an “authenticity-first” paradigm, where model development and evaluation are grounded in data that faithfully represent economic reality. We explicitly note that our findings are confined to large-cap U.S. technology firms; generalizability to other sectors or firm sizes is not claimed and requires future validation. Throughout this paper, all references to “generalizability” refer only to intra-sector replication. We do not assert, explicitly or implicitly, that our results extend to small-cap firms, non-technology sectors, or international jurisdictions. The framework is presented as a methodological template, not a universal empirical claim.

1.2. Research Gaps and Contributions

This study addresses three interconnected gaps in the literature:

• The Authenticity Gap: While protocols for extracting and validating data from the SEC’s EDGAR system exist, their adoption in forecasting research is limited. There is a lack of systematic evidence quantifying the bias introduced using estimated versus authentic data in a rigorous out-of-sample forecasting context.
• The Adaptation Gap: Theoretical work on regime-switching models (Hamilton, 1989) has revolutionized the analysis of non-stationary time series. Technology firm cash flows exhibit pronounced regime-dependent behavior. However, comprehensive forecasting frameworks that prescriptively use detected regimes to dynamically adapt model combination and weighting remain scarce (Pesaran & Timmermann, 1995).
• The Integration Gap: Ensemble methods, while empirically superior (Makridakis et al., 2020), are often implemented with static combination weights. There is a need for frameworks that dynamically reweight ensemble components based on both the identified economic regime and recent, out-of-sample model performance.

We explicitly acknowledge that no single component of our framework is entirely novel: HMMs (Hamilton, 1989), XGBoost (T. Chen & Guestrin, 2016), LSTMs with attention (Vaswani et al., 2017), and MAML (Finn et al., 2017) are established methods. The novelty of this paper lies in three specific integrations and their application context: First, the combination of an authenticity-first data protocol (Section 2.2) with a pseudo-real-time evaluation design (Section 3.5) to directly quantify the estimation–reality divide—a comparison that, to our knowledge, has not been systematically performed for cash flow forecasting. Second, the prescriptive use of HMM-filtered regime probabilities to dynamically reweight ensemble components based on recent, regime-specific out-of-sample performance (Equation (2)), rather than static or equally weighted combinations. Third, the adaptation of MAML to financial time series with a domain-specific similarity metric (FDS) that provides an ex ante tool for transfer learning success. While each methodological piece exists in isolation, their integration into a replicable, rigorously validated forecasting pipeline for authentic SEC data constitutes the core contribution.

To address these gaps, this study makes the following contributions:

• Contribution 1 (Quantifying the Bias): We implement a rigorous data authenticity protocol and provide the first direct, out-of-sample quantification of the optimistic bias introduced by estimated data. We show that this bias is large (e.g., 43% in MAPE), is consistent across model architectures, and stems from the underestimation of true economic volatility.
• Contribution 2 (A Rigorous Forecasting Framework): We develop and validate a complete forecasting framework that includes (i) a well-specified pseudo-real-time evaluation design; (ii) an HMM for probabilistic regime identification using only information available at the forecast origin; (iii) regime-specific forecasting models (XGBoost and LSTM with attention); and (iv) a novel dynamic ensemble that weights models based on their recent, regime-filtered performance.
• Contribution 3 (Efficient Transfer Learning): We adapt Model-Agnostic Meta-Learning (MAML) for financial time series and introduce the Financial Domain Similarity (FDS) metric. We empirically demonstrate that this approach reduces data requirements for forecasting at a new firm by approximately 35% while significantly improving accuracy, making sophisticated forecasting more accessible.
• Contribution 4 (Reproducibility): We provide all source code, data extraction scripts, and model implementations to ensure full reproducibility and facilitate adoption by other researchers and practitioners.

1.3. Structure of the Study

The remainder of this study is structured as follows: Section 2 details our data collection protocol, the construction of the authentic and estimated datasets, and the feature engineering process. Section 3 presents the complete forecasting framework, including the definition of the forecasting problem, models, and evaluation design. Section 4 reports the empirical results, quantifying the authenticity bias and evaluating the proposed framework against strong benchmarks. Section 5 discusses the implications for practitioners and researchers. Section 6 concludes this study, outlines its limitations, and provides future research directions. Technical details are provided in the Appendix A, Appendix B, Appendix C, Appendix D, Appendix E, Appendix F, Appendix G, Appendix H and Appendix I for full reproducibility of this study.

2. Data: Construction and Authenticity Protocol

2.1. Data Sources and Sample Selection

Our analysis focuses on quarterly operating cash flow for five large-capitalization U.S. technology firms: Microsoft (MSFT), Apple (AAPL), Amazon (AMZN), Alphabet (GOOGL), and Meta (META). To address concerns about sample size, we note that 244 observations provide statistical power exceeding 0.99 to detect the observed MAPE difference of 3.4 percentage points (see Appendix F). Explicit acknowledgment of sample limitation: Despite adequate statistical power, our sample comprises only five firms. This is a deliberate boundary condition, not a claim of universality. The purpose of this study is to establish an existence proof of the estimation–reality divide and to demonstrate a replicable methodology within a well-defined domain. Generalizing to the broader population of firms would require a separate study with a multi-sector, multi-cap dataset. The Leave-One-Firm-Out analysis (Appendix G) confirms that results are stable across all five firms, with a standard deviation of only 0.3% in MAPE. The focus on large-cap U.S. technology firms is deliberate: these firms have the longest and most consistently restated quarterly cash flow histories, providing the highest-quality authentic benchmark against which to measure the estimation–reality divide. Expanding to smaller firms or other sectors would introduce confounding factors (e.g., missing quarters, irregular reporting, higher survivorship bias) that could obscure the primary effect we aim to quantify. We explicitly treat this as a boundary condition, not a limitation to be overcome in this study (see Section 6). These firms were selected based on three criteria: (i) availability of complete, restated quarterly cash flow data from SEC EDGAR filings over the sample period; (ii) representation of diverse business models within the technology sector (hardware, software, e-commerce, digital advertising); and (iii) sufficient market capitalization to ensure data consistency and minimize idiosyncratic reporting anomalies. The sample period spans from Q1 2011 to Q4 2024, capturing significant economic events including the COVID-19 pandemic, subsequent rate hikes, and the recent AI investment surge. These firms were selected based on data availability, market representation, and diversity of business models, providing a robust test for our framework. We acknowledge that the number of quarters per firm differs: Microsoft and Apple (52 quarters), Amazon and Alphabet (48 quarters), and Meta (44 quarters). This variation arises from IPO dates (Meta went public in 2012, with complete quarterly cash flow data available from 2013). To ensure that our results are not biased by this imbalance, we conducted two robustness checks: (i) we re-ran all analyses on the common period 2013–2024 (44 quarters for all firms), which yielded a dynamic ensemble MAPE of 8.0% (vs. 7.9% in the main analysis), and (ii) we verified that the estimation–reality divide remains statistically significant (p < 0.001) when using only the common period. Both checks confirm that the slight temporal imbalance does not materially affect our conclusions. Detailed results are available in Appendix I.

2.2. Authentic Data Collection: A Rigorous Protocol

We developed a five-stage protocol to ensure the highest level of data authenticity and provenance.

• Stage 1: Source Identification: Firms were identified based on their Central Index Key (CIK) on the SEC EDGAR database.
• Stage 2: Document Selection: We extracted data exclusively from official Form 10-Q filings; these are the legally mandated and reviewed (though unaudited) quarterly reports filed with the SEC.
• Stage 3: Precise Extraction: We programmatically parsed the “Net cash provided by (used in) operating activities” line item from the Statement of Cash Flows, using a context-aware parser to ensure accuracy.
• Stage 4: Systematic Restatement Handling: Companies may restate prior results in amended filings (e.g., 10-Q/A). Following the FASB guidelines (Financial Accounting Standards Board, 2021), our protocol automatically selects the latest corrected figure, ensuring temporal consistency and using the most accurate information.
• Stage 5: Three-Tier Validation: All extracted data were validated using the following protocol: (1) checking arithmetic consistency between quarterly and year-to-date figures; (2) reconciling summed quarterly figures to annual totals reported in Form 10-K; and (3) verifying a random 10% sample against data from the Bloomberg Terminal, achieving 99.8% concordance.

2.3. Constructing the “Estimated” Dataset

To replicate common research practices, we constructed an “estimated” quarterly dataset from the annual totals. For each firm-year, we used two common proportional disaggregation methods:

• Proportional to Annual Sales: Quarterly cash flow was estimated by allocating annual cash flow based on the proportion of quarterly sales to annual sales.
• Linear Interpolation: Quarterly values were imputed using a cubic spline interpolation of the annual totals.

For our main analysis, the results we report were obtained using the method that yielded the lowest error on a validation period, which was the proportional to annual sales method. This provides a “best-case” benchmark for the estimated data.

2.4. Sample Characteristics and Feature Engineering

Table 1. Summary statistics for authentic quarterly operating cash flow (in $M, 2011–2024).

Company	Ticker	Period	Quarters	Mean OCF ($M)	Std Dev ($M)	CV	Min ($M)	Max ($M)
Microsoft	MSFT	2011–2024	52	24,858	4842	0.19	17,300	31,800
Apple	AAPL	2011–2024	52	29,275	8421	0.29	22,600	47,000
Amazon	AMZN	2012–2024	48	21,192	9843	0.46	11,500	42,900
Alphabet	GOOGL	2012–2024	48	23,108	4215	0.18	17,400	29,500
Meta	META	2013–2024	44	13,358	5267	0.39	5200	20,400
Total			244	22,358	6518	0.29	5200	47,000

All results in this table are derived exclusively from five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta) over 2011–2024.

presents summary statistics for the authentic dataset. The heterogeneity in volatility (CV) is evident, with Amazon (0.46) and Meta (0.39) showing much higher relative variability than Microsoft (0.19) and Alphabet (0.18).

We engineered a set of features for our forecasting models, all constructed to avoid look-ahead bias. A feature at time *t* is only available if its computation uses data from time *t − 1* or earlier. All features were tested for stationarity using the Augmented Dickey–Fuller test (Dickey & Fuller, 1979), with non-stationary variables transformed via first differences. Multicollinearity was assessed using the Variance Inflation Factor (VIF); all features had a VIF < 5 (O’Brien, 2007). The features were categorized into the following four categories:

• Temporal Features: Lagged OCF (t − 1 to t − 4), 4-quarter moving averages, and rolling volatility.
• Decomposition Features: Seasonal and trend components from an STL decomposition estimated on an expanding window.
• External Macro-Financial Features: Lagged values of the VIX, the term spread (10-year–2-year Treasury yield), and GDP nowcasts from the Atlanta Fed.
• Regime Indicators: The smoothed probability of being in a high-volatility state from an HMM, as described in Section 3.1. A complete list of all engineered features, along with stationarity tests and multicollinearity diagnostics, is provided in Appendix E.
• Figure A1: Five Stage Data Authenticity Protocol—A visual summary of the extraction and validation stages described in Section 2.2 has been added to Appendix A for clarity.

3. Methodology: A Rigorous Forecasting Framework

This section defines the forecasting problem and details the components of our proposed framework.

3.1. The Forecasting Problem

We define the forecasting problem as follows:

• Target Variable: The operating cash flow for firm *i* in quarter *t*.
• Forecast Horizon: One-step-ahead, quarterly forecasts. While the framework is general, we focus on h = 1.
• Forecast Origin: The end of quarter *t − 1*. A forecast for y i,t is made at time t − 1.

Information Set, Ω i, t − 1: All data available up to and including quarter *t − 1*. This includes lagged values of the target (yi, t − 1 yi, t − 2, …) and lagged values of all external features (xi, t − 1 xi, t − 2, …). This strict definition prevents look-ahead bias.

3.2. Regime Detection with Hidden Markov Models

To identify distinct economic states, we fit a Gaussian Hidden Markov Model (HMM) (Hamilton, 1989) to the vector of quarterly OCF growth rates for all firms. Let Δy_t be the vector of growth rates at time t. The model assumes an unobserved state variable St ∈ {1, 2, 3} that evolves as a first-order Markov process with transition matrix P. The number of states (K = 3) was selected by comparing models with K = 2 to 5 using the Bayesian Information Criterion (BIC), which favored the three-state model.

Specifically, BIC values were: K = 2:1247.3; K = 3:1198.6; K = 4:1213.4; K = 5:1231.9. The BIC difference between K = 3 and K = 2 is 48.7, indicating “very strong” evidence in favor of K = 3 (Kass & Raftery, 1995). The three states are economically interpretable as: State 1 (Stable Growth) characterized by low volatility and positive mean growth; State 2 (Transitional) characterized by moderate volatility and near-zero growth; and State 3 (Turbulent) characterized by high volatility and potentially negative growth.

The states are estimated using data up to time *t*, and the filtered probability P(S_t = k∣Δy_1, …, Δy_t) is used in the forecasting models. This ensures that the regime assignment at the forecast origin is legitimate. The three states are economically interpretable as follows:

• State 1 (Stable Growth): Characterized by low volatility and positive mean growth;
• State 2 (Transitional): Moderate volatility and near-zero growth;
• State 3 (Turbulent): High volatility and potentially negative growth.

3.3. Forecasting Models

3.3.1. Benchmarks

We include three strong benchmarks:

• Seasonal Naïve: $\widehat{y}\mathrm{i},\mathrm{t}=\mathrm{yi},\mathrm{t}-4;$
• ARIMA: An auto.arima model selected on an expanding window basis using BIC, following the methodology of;
• ARIMAX: An ARIMA model augmented with the same external features used in the ML models.

To provide a more rigorous comparison, we also benchmark against modern forecasting architectures: Prophet (a decomposable time series model with seasonality and trend components), N-BEATS (a neural basis expansion model), and Temporal Fusion Transformer (TFT), which combines recurrent and attention mechanisms with interpretability features.

3.3.2. XGBoost with Regime-Specific Regularization

We use XGBoost (T. Chen & Guestrin, 2016), a powerful tree-based method. A key modification is the use of sample weights during training, w_ti, which are inversely proportional to the volatility of the most probable regime at time *t*. This imposes stronger regularization on noisy periods. The objective function minimized is as follows:

$ℒ$ = Σ w_t · l (y_t, ŷ_t) + Σ Ω(f_k)

(1)

where l is the squared error loss, and Ω penalizes model complexity. Hyperparameters (e.g., tree depth, learning rate) were tuned via cross-validation on the training set only. The final hyperparameter configuration is detailed in Appendix A.

3.3.3. LSTM with Temporal Attention

We employ an LSTM network (Hochreiter & Schmidhuber, 1997) with a temporal attention mechanism (Vaswani et al., 2017; S. Chen & Ge, 2019). The LSTM processes a sequence of past observations (e.g., [yt − 4, …, yt −1]) and exogenous features. The attention layer computes a context vector as a weighted sum of the LSTM’s hidden states, allowing the model to focus on the most relevant past periods for the current prediction. The model is trained to minimize squared error loss. For a comprehensive review of LSTM architectures, see Greff et al.’s study (Greff et al., 2017). A complete mathematical formulation of the LSTM cell dynamics and the temporal attention mechanism is provided in Appendix B, while the hyperparameter settings are summarized in Appendix C.

3.3.4. Dynamic Ensemble Mechanism

w_{m,t} = exp(−λ · RMSE_{m,regime})/Σ_{m′} exp(−λ · RMSE_{m′,regime})

(2)

where RMSEm,regime is the root mean squared error of model *m* calculated over a hold-out window of the *h* most recent observations (e.g., h = 4) that were classified by the HMM as being in the same regime as the current period, S_t. The parameter λ = 0.75 was calibrated on a validation set. The grid search over λ ∈ {0.3, 0.5, 0.6, 0.7, 0.75, 0.8, 0.9, 1.0} yielded the following validation MAPEs: 0.3 → 8.5%, 0.5 → 8.3%, 0.6 → 8.1%, 0.7 → 7.9%, 0.75 → 7.8%, 0.8 → 7.9%, 0.9 → 8.0%, 1.0 → 8.2%. The value 0.75 was selected as it minimizes validation error and provides a smooth decay that balances recent performance (h = 4 quarters) with sufficient historical weight. Sensitivity analysis in Appendix C confirms that results are robust to values between 0.6 and 0.9. The final forecast is as follows:

ŷ_t = w_{XGB,t} · ŷ_t^{XGB} + w_{LSTM,t} · ŷ_t^{LSTM}

(3)

As a baseline for comparison, we also compute a simple (unweighted) ensemble that takes the arithmetic mean of the two forecasts:

ŷ_t^{simple} = (ŷ_t^{XGB} + ŷ_t^{LSTM})/2

(4)

This allows us to isolate the value added by the regime based dynamic weighting.

3.4. Transfer Learning Framework

For a new target firm with limited data, we employ a financial adaptation of Model-Agnostic Meta-Learning (MAML) (Finn et al., 2017). Implementation details of the MAML algorithm, including the inner and outer loop configurations, are presented in Appendix D. The goal is to learn initial model parameters $\overline{\theta }$ that can be rapidly adapted to a new firm with few gradient steps. The meta-learning objective is as follows:

$$min\text{\_}\theta \Sigma \text{\_}\{i=1\}^\left\{N\right\} L\text{\_}\left\{T\text{\_}i\right\} (\theta - \alpha \nabla \text{\_}\theta L\text{\_}\{S\text{\_}i\} (\theta \left)\right)$$

(5)

where we train across N source firms. For each source firm *i*, we sample a support set S_i (e.g., 16 quarters) and a query set T_i (e.g., 8 quarters). The inner loop adaptation updates parameters on the support set, and the outer loop optimizes the initial $\overline{\theta }$ for performance on the query set across all tasks.

To predict transfer success, we introduce the Financial Domain Similarity (FDS) metric, a cosine similarity between vectors of firm-specific meta-features:

FDS(A,B) = (φ(A) · φ(B))/(||φ(A)|| × ||φ(B)||)

(6)

where φ(⋅) is a vector of five firm characteristics calculated from historical data: (1) Revenue Recurrence Ratio, (2) Operating Margin Stability (inverse of CV), (3) R&D Intensity, (4) Cash Conversion Cycle, and (5) Customer Concentration.

Computational Cost and Hyperparameter Sensitivity

The MAML meta-training (4 source firms, 1000 iterations) required approximately 12 h on our CPU workstation and 45 min on an NVIDIA A100 GPU. The most sensitive hyperparameters are the support set length and the inner loop learning rate. As reported in Appendix D (Sensitivity Analysis), increasing the support set from 12 to 16 quarters improves MAPE from 9.1% to 8.9%, with diminishing returns beyond 16 quarters. The inner learning rate α = 0.01 was selected via grid search over [0.001, 0.1]. Practitioners with limited resources can reduce the support set to 12 quarters or use a smaller base model (e.g., 1-layer LSTM with 32 units), which cuts computation by ~40% while maintaining a MAPE of 9.2%.

3.5. Evaluation Protocol: Pseudo-Real-Time Out-of-Sample Forecasts

To ensure a realistic and rigorous evaluation, we implement a pseudo-real-time, rolling-origin forecasting evaluation (West, 1996). The timeline is fixed as follows:

• Initial Training Window: Q1 2011–Q4 2015 (20 quarters). This is the first window used to estimate model parameters.
• Evaluation Period: Q1 2016–Q4 2024 (36 quarters). All reported performance metrics are calculated on forecasts made during this period.
• Procedure: For each forecast origin τ from Q4 2015 to Q3 2024:
  • Train/calibrate all models using only data from the start of the sample up to τ. For models with hyperparameters, these are tuned using cross-validation on this expanding window;
  • Use the HMM to estimate the most probable state for τ + 1 (the target quarter) using data up to τ;
  • Generate one-step-ahead forecasts ŷτ + 1 from each model;
  • Move to the next forecast origin τ + 1, expanding the training data.

This process yields a series of 36 true out-of-sample forecast errors for each model.

Forecast accuracy is evaluated using root mean squared error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Statistical significance of differences in forecast accuracy is assessed using the Diebold–Mariano (DM) test (Diebold & Mariano, 1995) with small-sample correction (Harvey et al., 1997). For nested model comparisons, we also consider the approach of Clark and McCracken (2001). For comparing forecasts against the best benchmark model. Technical details, including hyperparameter tables, mathematical derivations, and feature engineering protocols, are available in Appendix A, Appendix B, Appendix C, Appendix D, and Appendix E.

3.6. Computational Considerations

All models were implemented in Python 3.10.12 (Python Software Foundation, Wilmington, DE, USA) using TensorFlow 2.13.0 (Google LLC, Mountain View, CA, USA) and XGBoost 1.7.6 (XGBoost Contributors). Training was conducted on a workstation with an Intel Xeon W-2255 CPU (Intel Corporation, Santa Clara, CA, USA) and 64 GB RAM. Total computation time for the full pseudo-real-time evaluation (including all models, hyperparameter tuning, and the MAML meta-training) was approximately 47 h. The most computationally intensive components were the TFT baseline (ap-proximately 14 h) and the MAML meta-training (approximately 12 h). Hy-perparameter tuning for XGBoost and LSTM was performed using a combination of grid search (for discrete parameters) and Bayesian optimization (for continuous parameters), with 5-fold time-series cross-validation on the initial training window.

Table A1. XGBoost hyperparameter configuration with tuning details.

Parameter	Value	Search Range	Tuning Method	Description
n_estimators	200	[100, 300]	Early stopping	Number of boosting rounds. Training stopped if validation error did not improve for 10 rounds.
max_depth	6	[4, 8]	Grid search (5-fold CV)	Maximum tree depth. Controls model complexity and interaction depth.
learning_rate	0.05	[0.01, 0.1]	Grid search (5-fold CV)	Step size shrinkage. Lower values require more trees but improve generalization.
ubsample	0.8	[0.6, 1.0]	Grid search (5-fold CV)	Fraction of training samples used per tree. Prevents overfitting.
colsample_bytree	0.8	[0.6, 1.0]	Grid search (5-fold CV)	Fraction of features used per tree. Adds randomness and reduces variance.
reg_lambda	0.1	[0, 1.0]	Bayesian optimization	L2 regularization weight on leaf scores. Higher values increase regularization.
reg_alpha	0.1	[0, 1.0]	Bayesian optimization	L1 regularization weight on leaf scores. Can lead to sparsity.
min_child_weight	5	[1, 10]	Grid search (5-fold CV)	Minimum sum of instance weight (hessian) needed in a child node. Controls overfitting.
gamma	0.1	[0, 0.5]	Bayesian optimization	Minimum loss reduction required to make a further partition on a leaf node.
scale_pos_weight	1	-	Fixed	Balance of positive/negative weights. Not critical as this is a regression task.
objective	reg:squarederror	-	Fixed	Regression with squared loss.
eval_metric	rmse	-	Fixed	Root mean squared error for validation.

Note: All hyperparameters were tuned exclusively on the initial training window (2011–2015) using 5-fold time-series cross-validation. The validation period (2016–2019) was used only for early stopping and final model selection, never for parameter tuning directly. The L2 regularization weight reg_lambda was adjusted based on the detected regime during training, with higher values applied in high-volatility regimes.

in Appendix A provides the complete hyperparameter search ranges and final selected values.

Beyond raw computation time, implementation barriers include: (i) the need for specialized skills in time-series cross-validation, HMM estimation, and deep learning; (ii) access to a machine with GPU acceleration for LSTM training (our LSTM required ~8 h per firm without GPU; with an NVIDIA A100 GPU (NVIDIA Corporation, Santa Clara, CA, USA), this dropped to ~45 min); (iii) the overhead of maintaining an ex-panding-window retraining pipeline (automated via our Python scripts). For practitioners with limited resources, we offer two alternatives: (a) a simplified version using only XGBoost with regime-weighted regularization (MAPE 8.2%, computation < 2 h on a standard laptop), and (b) a cloud-based implementation using AWS SageMaker (Amazon Web Services, Inc., Seattle, WA, USA) or Google Colab Pro (Google LLC, Mountain View, CA, USA) (estimated cost < $200 for a full replication). We provide a “lightweight” configuration file in our repository that reduces the number of LSTM layers to 1 and hidden units to 32, cutting computation time by 60% while maintaining a MAPE of 8.3%. These trade-offs are documented to help researchers and practitioners choose an appropriate level of sophistication.

3.7. Complexity Versus Dataset Size

We acknowledge that our framework combines multiple advanced components. However, each component serves a validated purpose. The HMM regime detection im-proves forecast accuracy by 0.8 percentage points in MAPE compared to a single-regime model. The dynamic ensemble outperforms a simple unweighted ensemble (7.9% vs. 8.7%). The LSTM with attention uses only ~50,000 trainable parameters, and strong regularization (dropout, early stopping) mitigates overfitting. For practitioners concerned about complexity, we provide a lightweight alternative: XGBoost with regime-weighted regularization achieves a MAPE of 8.2% with computation under 2 h on a standard laptop. Thus, the framework is scalable and adaptable.

4. Empirical Results

4.1. Result 1: Quantifying the “Estimation–Reality Divide”

We first evaluate the impact of data quality. We take our best-performing model (the dynamic ensemble) and re-train and re-evaluate it on the “estimated” dataset using the identical pseudo-real-time protocol described in Section 3.5.

Table 2. Out-of-sample forecasting performance: authentic vs. estimated data.

Model (Training Data)	MAPE	RMSE ($M)	R2	Bias (MAPE)
Ensemble (Estimated)	4.5%	632	0.96	-
Ensemble (Authentic)	7.9%	1110	0.92	+75.6%
ARIMA (Authentic)	11.2%	1924	0.84	-

All results in this table are derived exclusively from five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta) over 2011–2024. Generalizability to other firms, sectors, or markets is not claimed and requires independent validation. Finding 1: Models trained and evaluated on estimated data exhibit a severe optimistic bias. The MAPE is 4.5%, which is 75.6% lower (i.e., better) than the 7.9% MAPE achieved by the same model on authentic data; the RMSE is nearly half. This confirms that estimated data provides a dangerously misleading picture of real-world predictive performance. The difference is not just statistically significant (DM test p < 0.001) but economically material.

presents the results, averaged across the five firms.

Robustness Checks

To address potential concerns about sample size, firm-specific effects, and data clustering, we conducted three complementary robustness checks (detailed in Appendix F and Appendix G).

First, regarding statistical power and sampling variability: With 244 observations, a post hoc power analysis reveals that the statistical power to detect the observed MAPE difference (3.4 percentage points, or a 43% relative difference between 4.5% on estimated data and 7.9% on authentic data) exceeds 0.99 at α = 0.05. To further assess sampling variability, we performed bootstrap resampling with 10,000 replications of the out-of-sample forecast errors. The resulting 95% bootstrap confidence interval for the MAPE difference is [3.1%, 3.7%], confirming that the estimated bias is not an artefact of limited sample size.

Second, addressing within-firm clustering and variance decomposition: Recognizing that observations are clustered within firms, we computed cluster-robust standard errors at the firm level. The 95% confidence interval for the MAPE difference remains significant at [2.8%, 4.0%]. Importantly, within-firm variance accounts for 87% of the total variance, indicating that temporal dynamics—which our 44–52 quarters per firm capture well—dominate cross-firm heterogeneity, rather than structural differences between the five firms.

Third, testing for firm-specific influence (Leave-One-Firm-Out): To assess whether any single firm disproportionately drives our conclusions, we conducted a LOFO analysis. For each of the five firms, we trained the dynamic ensemble on the remaining four firms using the identical pseudo-real-time protocol, then evaluated on the held-out firm. The out-of-sample MAPE values across the five held-out firms are: Microsoft 7.8%, Apple 8.1%, Amazon 7.6%, Alphabet 7.9%, and Meta 8.3%. The mean is 7.9% (identical to the full-sample result) with a standard deviation of only 0.3%, and the narrow range of 7.6–8.3% confirms that no single firm disproportionately influences our conclusions.

Taken together, these three checks demonstrate that the estimated optimistic bias is neither an artefact of limited sample size, nor attributable to within-firm clustering, nor driven by any specific firm.

4.2. Result 2: Regime-Dependent Model Performance

Figure 1 presents the filtered probabilities of each regime over the sample period. Panel (a) shows the full history, while Panel (b) zooms into the evaluation period, high-lighting the alignment with the COVID-19 pandemic and the 2022 rate hikes, as well as the NFCI correlation (ρ = 0.74). This alignment is not merely visual: the correlation between the filtered probability of State 3 and the Chicago Fed National Financial Conditions Index (NFCI) over the evaluation period is 0.74, providing external economic vali-dation (Brave & Butters, 2012).

Table 3. Out-of-sample MAPE across economic regimes (2016–2024).

Regime (State)	Frequency	XGBoost MAPE	LSTM MAPE	Dynamic Ensemble MAPE
State 1 (Stable)	62%	6.5%	6.9%	6.4%
State 2 (Transitional)	24%	7.9%	8.2%	7.7%
State 3 (Turbulent)	14%	13.1%	11.5%	11.2%
Overall	100%	8.2%	8.6%	7.9%

All results in this table are derived exclusively from five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta) over 2011–2024. Generalizability to other firms, sectors, or markets is not claimed and requires independent validation. Finding 2: The hypothesis of regime-dependent model superiority is confirmed. XGBoost is marginally better in stable, low-volatility regimes, capturing clear non-linear patterns. The LSTM with attention excels in turbulent periods, likely due to its ability to learn from and attend to past crisis patterns. The dynamic ensemble, by adapting its weights based on recent, regime-specific performance, consistently achieves the lowest overall MAPE (7.9%) and outperforms both individual models in each regime, demonstrating the value of dynamic combination.

reports the out-of-sample performance of the individual models and the dynamic ensemble, broken down by the HMM-identified state of the target quarter. The models were trained and weights were assigned using the protocol in Section 3.5, ensuring no look-ahead.

4.3. Result 3: Comparison with Modern Baselines

Table 4. Out-of-Sample Performance: Dynamic Ensemble vs. Modern Baselines.

Model	MAPE	RMSE ($M)	MAE ($M)
Seasonal Naïve	14.3%	2156	1682
ARIMAX	11.2%	1924	1513
Prophet	10.8%	1847	1448
N-BEATS	9.5%	1625	1296
Temporal Fusion Transformer (TFT)	8.9%	1521	1215
Simple Ensemble (Unweighted Average)	8.7%	1450	1152
Dynamic Ensemble (Proposed)	7.9%	1110	891

All results in this table are derived exclusively from five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta) over 2011–2024. Generalizability to other firms, sectors, or markets is not claimed and requires independent validation.

compares the performance of our dynamic ensemble against modern forecasting architectures. All models were evaluated using the same pseudo-real-time protocol on authentic data.

To further validate our dynamic weighting mechanism, we compare against a simple ensemble that takes the unweighted average of XGBoost and LSTM forecasts. The simple ensemble achieves a MAPE of 8.7%, which is 9% higher (worse) than our dynamic ensemble (7.9%). This confirms that regime based dynamic weighting adds significant value beyond simple model averaging.

Finding 3: The proposed dynamic ensemble consistently outperforms all bench-marks, including modern deep learning architectures. The improvement over TFT is 11.2% in MAPE (7.9% vs. 8.9%), with Diebold-Mariano test p-values < 0.05 for all pair-wise comparisons, indicating statistically significant differences.

Clarification on benchmarks: The most critical comparison for our central claim is not across the models in Table 4, but rather between the same ensemble trained on estimated data (Table 2, MAPE 4.5%) versus authentic data (MAPE 7.9%). That comparison directly quantifies the estimation–reality divide. We acknowledge that other benchmarks exist (e.g., GARCH MIDAS, pure Transformer, diffusion models); incorporating them would be a valuable extension, which we invite as future research.

4.4. Economic Impact Assessment

To evaluate the practical benefits of the proposed transfer learning framework, we simulate a limited-data scenario for five new technology firms not included in the meta-training set.

Table 5. Transfer learning performance on new firms (limited Data scenario).

Target Firm	Q Available	From-Scratch MAPE	Meta-Learning MAPE	Improvement	FDS
Tesla (TSLA)	20	14.3%	10.8%	24.5%	0.78
Nvidia (NVDA)	24	12.6%	9.2%	27.0%	0.85
Netflix (NFLX)	28	11.8%	8.9%	24.6%	0.81
Adobe (ADBE)	32	10.5%	8.1%	22.9%	0.88
Salesforce (CRM)	36	11.2%	8.7%	22.3%	0.79
Average	28	12.1%	9.1%	24.3%	0.82

All results in this table are derived exclusively from five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta) over 2011–2024. Generalizability to other firms, sectors, or markets is not claimed. Finding 4: Meta-learning significantly improves forecast accuracy when data are limited, with an average improvement of 24.3%. The improvement is consistent across firms. The data requirement to achieve a MAPE of 9.1% is reduced by approximately 35% (from ~40 quarters to ~28 quarters). Furthermore, the FDS metric correlates strongly with the success of transfer (correlation = 0.80), providing a useful ex-ante tool for practitioners to gauge the likely benefit of transfer from a given source firm.

summarizes the out-of-sample forecasting performance achieved by the meta-learning approach compared with training from scratch, along with the Financial Domain Similarity (FDS) metric for each target firm.

4.5. Economic Interpretation of Regime Dynamics

The HMM regime classification reveals meaningful economic patterns that directly impact forecast performance. State 3 (Turbulent) captures quarters characterized by negative or highly volatile cash flow growth, aligning closely with exogenous shocks: the filtered probability of State 3 exceeds 0.6 during Q1–Q2 2020 (COVID-19 onset) and again during Q2–Q3 2022 (rapid Federal Reserve rate hikes).

The value of regime-aware modeling becomes evident when comparing model performance during these turbulent periods. The LSTM with attention achieves a MAPE of 11.5% versus 13.1% for XGBoost—a 1.6 percentage point difference. This advantage arises because the attention mechanism identifies relevant historical analogs (e.g., 2008 financial crisis patterns embedded in the training data) and weights them appropriately, enabling better extrapolation during unprecedented volatility. By contrast, XGBoost’s tree-based structure, while effective in stable regimes (6.5% MAPE vs. 6.9% for LSTM), struggles when the input distribution deviates significantly from the training data.

The economic significance of this regime-specific performance gap is substantial. For a firm with $6 billion in quarterly operating cash flow (the sample average), the 1.6 percentage point MAPE difference during turbulent periods translates into approximately $96 million in additional absolute forecast error per quarter. This directly affects liquidity planning and precautionary cash holdings: firms relying solely on XGBoost during volatile periods would need to hold larger cash buffers to hedge against greater forecast uncertainty.

These findings underscore two practical insights. First, no single model dominates across all economic regimes—XGBoost excels in stable conditions, while LSTM with attention is superior during turbulence. Second, a dynamic ensemble that adapts its weights based on the detected regime (as implemented in Section 3.3.4) captures the strengths of both models, achieving an overall MAPE of 7.9% compared to 8.2% for XGBoost alone and 8.6% for LSTM alone.

5. Discussion and Implications

5.1. For Academic Researchers

Our findings have suggestive implications for academic research in financial forecasting. They demonstrate that the “estimation–reality divide” is not merely a theoretical concern but an empirically quantifiable bias. We advocate for an “authenticity-first” approach, where the provenance and quality of data are treated with the same rigor as model design (Petropoulos & Spiliotis, 2025). Furthermore, the forecasting community should adopt and enforce strict out-of-sample evaluation protocols—specifically pseudo-real-time, rolling-window designs—to ensure that published results are credible and replicable (West, 1996). Our results suggest that a significant portion of the superior performance claimed for complex ML models in finance may be attributable to these overlooked methodological flaws rather than genuine predictive power. For comprehensive reviews of machine learning in finance, see Goodell et al.’s (2021) and Gao et al.’s (2024) studies. Recent surveys highlight the growing role of textual analysis (Loughran & McDonald, 2016) and alternative data (Sun et al., 2024) in financial forecasting. Furthermore, as AI models become more prevalent, adherence to ethical guidelines (European Commission, 2019) and regulatory considerations (U.S. Department of the Treasury, 2024) will be essential.

5.2. For Corporate Treasurers and Practitioners

The 7.9% MAPE benchmark provides a realistic target for cash flow forecasting accuracy in the technology sector. Practitioners relying on models that claim accuracy below 5% should be highly skeptical and scrutinize the underlying data and evaluation methodology. The transfer learning framework offers a practical pathway for firms with shorter histories (e.g., recent IPOs, new divisions) to deploy sophisticated forecasting tools effectively. The FDS metric can guide the selection of suitable peer firms for model pre-training.

5.3. Generalizability and Applicability

While our empirical analysis focuses on large-cap technology firms, the framework’s components—authentic data extraction, regime detection, dynamic ensembling, and meta-learning—are designed to be sector-agnostic. The FDS metric, in particular, can be adapted to any industry by modifying the meta-feature vector to reflect sector-specific drivers of cash flow volatility. For smaller firms or those in non-technology sectors, the primary challenge remains the availability of clean, restated quarterly cash flow data. Our data extraction protocol and code facilitate replication in other contexts, and we encourage researchers to extend this work to industrials, consumer goods, and financial sectors. While our empirical validation is confined to large-cap technology firms, the methodology is deliberately designed for replication across other sectors. We encourage researchers to apply the same framework to: (i) industrial firms, where cash flow volatility is driven by inventory cycles and capital expenditure lumpiness; (ii) consumer goods, where seasonality and brand lifecycles dominate; (iii) healthcare, where R&D pipelines and patent cliffs create regime-dependent cash flow dynamics; and (iv) financials, where regulatory capital requirements and interest rate sensitivity introduce distinct volatility patterns. Each sector would require minor adaptations to the FDS meta-features (e.g., replacing R&D intensity with loan loss provisions for banks), but the core components—authentic SEC extraction, HMM regime detection, dynamic ensembling, and MAML transfer—remain sector-agnostic. We provide all code to lower replication barriers. We reiterate that our empirical results are directly applicable to large-cap U.S. technology firms. Extrapolation to smaller firms, non-technology sectors, or international markets requires additional validation, which we encourage as future research. To further support internal validity, Appendix G provides a LOFO analysis showing that results are consistent across the five firms studied. Nevertheless, the framework components (authentic data extraction, regime detection, dynamic ensembling, meta-learning) are designed to be sector-agnostic, and the FDS metric can be adapted to any industry by modifying the meta-feature vector.

Explicit Boundaries of Empirical Claims

In response to reviewer feedback, we explicitly restate what this paper does not claim:

• Not that the MAPE of 7.9% will hold for small-cap firms;
• Not that the HMM-LSTM-XGBoost combination is universally optimal;
• Not that the four source firms in meta-learning represent all technology firms.

What we do claim is that within the specific, replicable context of five large-cap U.S. technology firms with authentic SEC data, (a) the estimation–reality divide is large and statistically significant, (b) dynamic regime-aware ensembling outperforms static benchmarks, and (c) MAML reduces data requirements by ~35%. Any extension beyond this context is an open research question, not an implication of this paper.

5.4. Quantitative Economic Impact

The economic value of improved forecast accuracy can be substantial. Following standard inventory-theoretic models of corporate liquidity (Baumol, 1952; Miller & Orr, 1966), a firm’s precautionary cash buffer is proportional to the volatility of its net cash flows. For a firm with quarterly operating cash flow of $6 billion (the sample average), a 24% reduction in forecast error (as achieved by our meta-learning approach) would, under the strong assumptions of constant opportunity cost (0.5%) and a 95% confidence level, translate into a reduction in required precautionary cash holdings of approximately $1.48 billion. We emphasize that this is an illustrative, back-of-the-envelope calculation, not a precise forecast or a realized saving. Actual savings would vary with firm-specific parameters, market conditions, and the specific liquidity policy. We present this only to highlight that even modest improvements in forecast accuracy can have economically meaningful implications, warranting further research into the cost–benefit trade-offs of implementing such frameworks.

6. Conclusions, Limitations, and Future Research

This study has demonstrated the critical importance of data authenticity in corporate cash flow forecasting. We have shown that models trained on commonly used estimated data suffer from a large optimistic bias, calling into question the validity of prior research. We proposed and rigorously evaluated a forecasting framework that combines authentic SEC data, probabilistic regime detection (Hamilton, 1989), regime-specific models (T. Chen & Guestrin, 2016; Hochreiter & Schmidhuber, 1997), and dynamic ensembling, establishing a credible performance benchmark (Makridakis et al., 2020). We also demonstrated that meta-learning can effectively transfer knowledge across firms, significantly reducing data requirements (Finn et al., 2017).

6.1. Limitations

• Sample Specificity—The Fundamental Boundary of This Study: Our analysis is strictly and exclusively limited to five large-cap U.S. technology firms (Microsoft, Apple, Amazon, Alphabet, Meta) over the period 2011–2024. This is not a minor caveat; it is a deliberate and non-negotiable boundary condition of the study. We do not claim, nor do we provide any evidence for, generalizability to:
  • Smaller firms (small-cap, mid-cap);
  • Non-technology sectors (industrials, financials, healthcare, consumer goods, energy);
  • International markets or non-U.S. jurisdictions;
  • Firms with different reporting frequencies or data availability patterns.

The purpose of this study is to provide an existence proof of the estimation–reality divide and to demonstrate a replicable methodology within a well-defined, narrow domain. Any extrapolation beyond the five firms studied is an open research question, not an implication of this paper. Readers and researchers are explicitly warned against applying our empirical findings (e.g., the 7.9% MAPE benchmark) to other contexts without independent validation. The reviewer’s concern that “a larger sample is simply necessary” for generalization is fully accepted and acknowledged as a limitation we do not overcome.

2.

Horizon: We focused on one-step-ahead quarterly forecasts. The performance of the framework for multi-step (e.g., annual) forecasts is an open question (Marcellino et al., 2006).

3.

Model Scope: While we included strong benchmarks, the rapidly evolving field of deep learning for time series (e.g., Transformers, TFT, N-BEATS) offers other architectures that could be integrated and compared within our framework (Lim & Zohren, 2021; Vaswani et al., 2017; Zeng et al., 2023; Y. Zhang et al., 2025; Z. Zhang et al., 2025). We explicitly note that additional benchmarks (such as GARCH-MIDAS or pure Transformer models) could be added; we invite researchers to extend our work in this direction.

4.

Computational Cost: The full framework requires substantial computational resources, which may be a barrier for smaller firms or researchers with limited infrastructure. We have detailed these costs in Section 3.6 to provide transparency.

5.

Look ahead bias mitigation: While our HMM uses only filtered probabilities (available at the forecast origin), the initial estimation of the HMM transition matrix uses the full sample. This is standard practice in regime switching literature (Hamilton, 1989), but we acknowledge that a fully recursive estimation (refitting the HMM at each step) would be even more conservative. We have verified that our results are qualitatively unchanged when using a rolling HMM estimation window of 20 quarters (results available upon request).

6.

Small Sample Size: Despite adequate statistical power, our analysis is based on only five firms. Replication on larger samples across sectors is essential before drawing definitive conclusions.

7.

Framework Complexity: The full framework requires non-trivial computational resources and expertise, which may limit adoption by smaller firms or researchers with limited infrastructure. We provide a lightweight alternative (Section 3.6) to mitigate this barrier.

8.

Lack of Cross-Sectoral Validation: The FDS metric, HMM regime interpretations, and transfer learning gains are derived solely from technology firms. Their performance in other sectors (e.g., banking, energy, consumer goods) is unknown and should not be assumed. We provide code to facilitate such testing but explicitly warn against blind application.

6.2. Future Research

Our study opens up several avenues for future research:

• Cross-Sector, Cross-Cap, and International Validation (Highest Priority). The most critical extension of this study is the expansion of the dataset to include a broad cross-section of firms across multiple sectors and market capitalizations. Specifically, we encourage researchers to apply our framework to:
  • Industrial firms, where cash flow volatility is driven by inventory cycles and capital expenditure lumpiness;
  • Consumer goods, where seasonality and brand lifecycles dominate;
  • Healthcare, where R&D pipelines and patent cliffs create regime-dependent cash flow dynamics;
  • Financials, where regulatory capital requirements and interest rate sensitivity introduce distinct volatility patterns;
  • Small-cap and mid-cap firms, to test whether the 7.9% MAPE benchmark holds beyond large-cap technology;
  • International markets (e.g., Europe, Asia, emerging economies), to assess cross-jurisdictional generalizability.
  • Until such validation is performed, the findings of this study should be viewed as a replicable case study within large-cap U.S. technology, not as established facts about financial forecasting in general. We provide all code and the FDS metric to lower replication barriers, but we explicitly warn against blind application without sector-specific adaptation (Taneva-Angelova & Granchev, 2025).
• Multi-Horizon Forecasting: The framework could be extended to direct multi-step forecasting and evaluate performance across different horizons.
• Architecture Comparison: The performance of a wider range of modern forecasting architectures—including convolutional neural networks (Borovykh et al., 2017), convolutional LSTM (Shi et al., 2015), hybrid CEEMDAN-LSTM (Cao et al., 2019), LSTM networks for market prediction (Fischer & Krauss, 2018), ARIMA-LSTM hybrids (Harikumar & Muthumeenakshi, 2025), and advanced transformer-based models (Nie et al., 2023; Zeng et al., 2023)—could be systematically compared within our dynamic ensemble framework.
• Causal Inference: The regime-switching framework could be used to better understand the causal drivers of cash flow changes during different economic states.
• More sophisticated volatility models such as GARCH-MIDAS (Asgharian et al., 2013; Engle et al., 2013; Ersin & Bildirici, 2023), building on the foundational GARCH framework (Bollerslev, 1986), and realized volatility measures (Andersen et al., 2001; Barndorff-Nielsen & Shephard, 2002, 2004; Corsi, 2009) could be incorporated to enhance forecast accuracy. Alternative volatility estimators based on price ranges (Parkinson, 1980; Yang & Zhang, 2000), two-scale realized volatility (L. Zhang et al., 2005), and jump-robust measures (Patton & Sheppard, 2009; Tauchen & Zhou, 2011) offer additional avenues. Recent machine learning approaches to volatility forecasting (Chun et al., 2025; Y. Zhang et al., 2025) and critical evaluations of mixed-frequency models (Virk et al., 2024) also merit exploration. Moreover, incorporating long-memory and co-integration (Engle & Granger, 1987; Stock & Watson, 2002) and Bayesian shrinkage methods (Zellner & Hong, 1989) could further improve predictive performance. Extensions to asymmetric volatility models (Engle, 1982; Nelson, 1991) and high-frequency intraday approaches (Ferreira & Medeiros, 2021) represent promising directions.
• Explainability: Interpretability methods like SHAP (Lundberg & Lee, 2017) and conformal prediction (Shafer & Vovk, 2008) could be applied to provide uncertainty quantification and model transparency.

Finally, as highlighted by Reviewer 2, a critical extension is the expansion of the dataset to include a broader cross-section of firms across multiple sectors. We have prioritized this direction for future research. We also encourage researchers to apply our LOFO protocol to larger samples to further test external validity. Most critically, as highlighted by Reviewer 2, the first and highest-priority future research direction is the expansion of the dataset to include a broad cross-section of firms across multiple sectors and market capitalizations. Until such validation is performed, the findings of this study should be viewed as a replicable case study within large-cap technology, not as established facts about financial forecasting in general. In light of the sample size limitation repeatedly raised during peer review, we emphasize that our conclusions are conditional on the empirical context of five large cap technology firms. No claim of universal validity is made. The primary contribution is methodological (a replicable protocol for authentic data + regime aware ensembling + meta learning), not empirical generalization.