Year: 2025
Type: article
Abstract: This study tests whether modern machine-learning models can generate economically meaningful alpha when forecasting daily U.S. large-cap returns under realistic trading frictions. Using 11 years of high-resolution data (2013-2023) on 423 liquid S&P-500 constituents, we benchmark six architectures OLS, ARIMA-GARCH, Random Forest, XGBoost, LSTM and the attention-based Temporal Fusion Transformer (TFT) within a rolling 1 260-day/252-day walk-forward protocol. TFT attains the highest directional accuracy (58.7 %, F1 = 0.57) and produces a post-cost Sharpe ratio of 1.31 versus 0.96 for XGBoost; an equal-weight ensemble of TFT, LSTM and XGBoost delivers 1.42 Sharpe with a maximum draw-down of –9.6 %. Ablation studies show technical indicators dominate predictive power, sentiment adds 0.08 Sharpe, and macro variables contribute 0.05. The edge persists at transaction costs up to 10 bps and across the COVID-19 and GFC regimes. While GPU inference and nightly retraining introduce operational overhead, the incremental 0.4 Sharpe remains economically large. Our open-source pipeline and dataset enable full replication and extension.
Authors: Sidrah Kaleem, Hafsa Rasheed, Anees Shahzad
Institutions: Rawalpindi Women University, Linnaeus University
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: bronze
OpenAlex ID: W4413942144