Introduction: Fundamental and trading data provide different perspectives on stock return predictability, and their combined predictive value may depend on the forecasting horizon and how the information is integrated. Accounting information is generally updated quarterly or annually, whereas trading data capture daily market developments. Valuation ratios involving market capitalization can also change daily. This study evaluates a two-stage residual learning framework for cross-sectional stock return prediction over horizons of 1, 3, 5, 10, and 15 trading days. Its performance is compared with fundamental-only, trading-only, and direct feature-combination models to examine whether sequential integration improves ranking quality and return separation.



Methods: The study examines manufacturing companies listed on the Tehran Stock Exchange. Financial statement information was obtained from Codal, and daily price and trading data were collected from TSETMC. Fundamental features included trailing-twelve-month earnings-to-market-capitalization and sales-to-market-capitalization ratios, quarterly earnings and sales growth, and changes in gross, operating, and net profit margins. These financial features were transformed into daily cross-sectional percentile ranks across the sample and within industries. Accounting information was incorporated from the day after its official publication. Trading features captured candlestick characteristics, historical returns, and measures related to risk and market movements.



All four designs used XGBoost regression and were evaluated through two successive rolling test windows. Training and validation observations whose target-return intervals crossed the respective period boundaries were excluded. In the two-stage framework, the first model predicted returns using fundamental features, and the second used trading features to learn residuals calculated from the first model’s predictions on the same training observations. Target returns were measured from the adjusted closing price on the prediction date to the adjusted opening price at the forecast horizon. Performance was evaluated using the mean daily Spearman Rank Information Coefficient (Rank IC) and the return spread between the highest- and lowest-scoring deciles, with common evaluation filters applied across models.



The sample comprised 274 manufacturing firms listed on the Tehran Stock Exchange and approximately 400,000 firm-day observations from January 2017 to June 2026. Approximately 70,000 observations were used in two consecutive test windows, July 2024–June 2025 and July 2025–June 2026, with the remaining observations used for training and validation.





Findings: The two-stage model achieved the highest mean Rank IC at all five forecasting horizons, with values of 0.32987, 0.04453, 0.03668, 0.05328, and 0.04400 at 1, 3, 5, 10, and 15 days, respectively. At the 1-day horizon, its performance was close to that of the trading-only and direct-combination models, which recorded IC values of 0.32856 and 0.32875. The advantage was more pronounced at 3 days: the two-stage model achieved an IC of 0.04453 and a spread of 0.483%, compared with 0.03103 and 0.332% for the trading-only model and 0.02252 and 0.153% for direct combination.

The two-stage model also recorded the largest decile spreads at the 1-, 3-, and 15-day horizons, at 1.847%, 0.483%, and 1.83%, respectively. Direct combination produced higher spreads at 5 and 10 days, while the two-stage model retained the highest mean IC. Across all five horizons, the two-stage model’s spread exceeded that of the trading-only model. The fundamental-only model’s IC increased from 0.00247 at 1 day to 0.04211 at 15 days, indicating greater standalone predictive value at the longer horizons examined.



Conclusions: The two-stage residual learning framework achieved the highest mean Rank IC across all five forecasting horizons and the largest decile spread at the 1-, 3-, and 15-day horizons. Its spread also exceeded that of the trading-only model at every horizon. These findings support the value of sequentially integrating fundamental and trading information to improve cross-sectional stock ranking. Direct combination produced the largest spreads at 5 and 10 days, indicating that the preferred integration method also depends on the forecasting horizon and evaluation objective.