Benchmarking Transformers and Baselines for Multi-Horizon Stock Return Prediction with Technical and Earnings Features
Abstract
We study daily multi-horizon stock return prediction on the Dow $30(h \in\{1,5,21\})$ using technical indicators and earnings features under a strict temporal split (train 2016-2019, val 2020, test 2021-2024). At h = 1, LSTM attains the lowest RMSE ($\mathbf{0. 0 1 6 2 2}$), narrowly ahead of GRU (0.01625), while Random Forest yields the highest DA $\boldsymbol{(} \mathbf{5 1. 9 \%} \boldsymbol{)}$. At $h=5$, Ridge achieves the best RMSE ($\mathbf{0. 0 3 6 5 1}$). At $h=21$, LSTM delivers both the lowest RMSE (0.05084) and highest DA (55.0%), edging TFT and GRU by $0.2-0.3 \%$ RMSE and 1 pp DA and outperforming tabular baselines by $\mathbf{3 1 - 3 8 \%}$ in RMSE. Adding earnings features increases DA by 1 pp at $h=21$, and DA rises to $\mathbf{6 1. 4 \%}$ within ±3-day earnings windows (vs. 53.9% otherwise). By regime, GRU performs best in the 2022 bear (RMSE 0.0636) and LSTM in the $2023-2024$ rally (RMSE 0.0475; DA 56.2%). Overall, simple sequence models match or surpass TFT on this constrained daily panel, with tabular methods competitive at shorter horizons.
Authors: Nelson Siu, Jonathan H. Chan