From Shots to Forecasts: One-Step-Ahead Prediction of Team-Level Expected Goals in Football

Expected goals (xG) describe football performance, but their value as time-series signals for forecasting future team performance is less established. Using source data covering 32,121 matches across 30 competitions, we evaluated one-step-ahead team-level xG forecasts under chronological within-panel splits and a calendar-cutoff sensitivity analysis. The primary target was focal-team xG (xGF); xG against (xGA) and xG difference (xGD) were derived from paired team forecasts. We compared rolling mean, ARIMA, linear mixed-effects, XGBoost, and Temporal ConvNet models. The mixed-effects model achieved the lowest test mean absolute error (MAE) for xGF (0.545 versus 0.591 for rolling mean) and xGD (0.790 versus 0.861), a moderate reduction of about 8%. XGBoost had a similar MAE and the highest xGD directional accuracy (66.35% versus 60.97%, a 5.38-percentage-point gain). The primary XGBoost context-only comparison jointly removed xG history, match week, and the number of previous matches without a statistically detectable MAE difference. Mixed-model history gains (0.0009 and 0.0020) were not significant after multiplicity adjustment. With similar input information, the Temporal ConvNet had comparable xGF MAE and a small xGD advantage over XGBoost. Validation-based smearing reduced negative bias and root mean squared error (RMSE) at a small MAE cost. Calendar evaluation showed history gains for both mixed-effects and XGBoost models in previously observed and new seasons. Non-significant differences under the primary retrospective within-panel split do not establish equivalence.