Beyond Data Leakage: Robust Pre-Release Forecasting of North American Video Game Sales

The video game market is large yet volatile, making reliable sales forecasts critical for publishers and retailers. Many prior approaches overlook temporal dependencies and risk data leakage, inflating accuracy. This study presents a leakage-free ensemble framework for forecasting North American video game sales. Time-aware features, specifically lagged “previous-title” sales and three-year rolling trends by genre, platform, and publisher, are engineered strictly from past information. Using VGChartz-derived historical data, Random Forest (RF), Gradient Boosting (GB), and histogram-based Gradient Boosting (HGB) are evaluated under a time-based train/test split and time-series cross-validation. Results show that the HGB consistently outperforms linear baselines, achieving a Coefficient of Determination (R2) of approximately 0.384, along with the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Feature-importance analyses highlight prior-title sales and publisher trends as dominant predictors. These findings demonstrate that avoiding leakage and incorporating temporal structure significantly improve forecast reliability, offering valuable insights for sales planning and strategy.