Multi-horizon stock price prediction with an informer-based framework leveraging inter-stock correlations and commodity inputs

A significant goal of stock price prediction research is to understand the relationships among stocks and how they move over time, enabling better stock price forecasts and portfolio risk assessment. This has the potential to impact asset allocation, price forecasts, and portfolio risk over time through changes in cross-asset movements. In this study, we develop an Informer-based framework to examine how inter-stock relations and exogenous commodity variables support multi-horizon stock price prediction. The proposed approach combines multi-source inputs with a deterministic sinusoidal encoding implemented in the encoder. This combination addresses the difficulty of correctly mapping intricate temporal and cross-asset relationships. The study examines whether time-varying and cross-asset relationships affect the price movements of individual stocks within an index when there is high inter-stock correlation. The framework utilizes both endogenous (historical prices and technical indicators) and exogenous commodity inputs (e.g., crude oil, gold, and silver prices). The experiments evaluated four feature selection approaches, including correlation-based filtering, k-best selection, shapley additive explanations (SHAP), and the full-feature set. Among these feature sets, the SHAP-based feature set achieved the best empirical performance under the evaluated configuration. The proposed model is evaluated using the time-series cross-validation method (k = 3, 5, 7) in four stocks, including INFY, TCS, HDFC Bank, and ICICI Bank, as well as the NIFTY 50 index. It achieves an average coefficient of determination (R \({^2}\) ) score of 0.987, along with a mean absolute percentage error (MAPE) of 3.098%, and a mean absolute scaled error (MASE) of 0.277, showing improved performance over the evaluated baseline models. A comparative ablation study was done comparing the performance of the proposed model against Transformer, Autoformer, and Reformer baselines, to examine the contribution of the proposed framework under the evaluated settings. We further evaluated the proposed framework using a profitability analysis with a rule-based trading simulation under simplified trading assumptions. Finally, two statistical tests, including the Wilcoxon signed-rank test and Diebold–Mariano (DM) test, were conducted to statistically compare the proposed model against benchmarks. The results suggest that related stock inputs provide useful information beyond individual stock histories and improve forecasting performance across multiple horizons.