Crude Oil Prices and S&P 500 Futures Returns: A Machine-Learning Approach Using Daily Data, 2016–2026
Elijah Kihooto Maringa *
University of Kigali, Kigali, Rwanda.
Ronald Emojong
Busitema University, Busitema, Uganda.
Nyirimana Jean Marie Vianney
University of Kigali, Kigali, Rwanda.
*Author to whom correspondence should be addressed.
Abstract
This study examines whether information from crude oil markets contributes to the prediction of daily E-mini S&P 500 futures returns using observations from 2016 to October 2026. The analysis incorporates Brent crude oil futures, West Texas Intermediate (WTI) crude oil futures and the CBOE Volatility Index alongside equity-market information. Pairwise correlation analysis, an ordinary least-squares specification with heteroskedasticity- and autocorrelation-consistent inference, and machine-learning feature-set comparisons were employed. S&P 500 futures returns showed positive correlations with Brent returns (0.20236) and WTI returns (0.12114), while Brent and WTI returns were moderately correlated (0.48743). In the reported regression, Brent and WTI coefficients were positive and statistically significant after controlling for VIX. However, the forecasting comparisons provided only limited evidence of incremental predictive value from oil-market information. The equity-plus-WTI specification produced the lowest reported root mean squared error of 1.04395, compared with 1.05080 for the equity-only benchmark, representing an improvement of approximately 0.65%. Its directional accuracy was 51.38%, compared with 50.28% for the equity-only specification. Other oil-containing feature sets did not consistently outperform the equity-only benchmark. Overall, the results suggest that WTI information may provide modest incremental predictive information within the evaluated sample, but the observed performance difference is small and does not establish a stable forecasting advantage or causal relationship. This manuscript contributes to the long-running debate over whether crude oil markets carry information that helps forecast equity returns. It tests that question out of sample on a recent and unusually eventful period, from 2016 to October 2026, which includes the 2020 oil-price collapse and later energy-market disruptions. By comparing equity-only and oil-augmented feature sets under both statistical and machine-learning approaches, it moves beyond in-sample correlation and shows that a significant regression coefficient does not necessarily translate into forecasting gain. The modest result, a roughly 0.65% reduction in RMSE and a directional accuracy of 51.38% against 50.28% for the equity-only benchmark, is informative in its own right. It points to limited incremental predictive value from oil information, which is a useful caution against overstating predictability in this literature. The study also gives researchers and practitioners a transparent, replicable benchmark with HAC-robust inference and explicit feature-set comparisons, which can be extended to other asset classes, volatility regimes and forecasting horizons.
Keywords: Crude oil, S&P 500 futures, machine learning, XGBoost, WTI, Brent, forecasting, financial return