Feature Selection in Bitcoin Price Forecasting: Temporal Persistence Across Walk-Forward Periods and Learning Models

Bitcoin price forecasting often uses broad information sets that combine market, on-chain, macroeconomic, financial, energy, and technical variables. This study examines whether predictor reduction improves one-day-ahead Bitcoin price forecasts within learning models and whether this effect persists across successive out-of-sample periods. Daily data from January 2015 to May 2024 are analyzed. A nine-variable specification is first selected from an eleven-variable information set using Drop-Column ranking and nested Top-K selection. The selected specification is evaluated with Random Forest (RF), XGBoost, and Long Short-Term Memory (LSTM) models using a fixed out-of-sample evaluation and ten successive 60-day walk-forward windows. Under observation-date alignment, the selected specification reduces RMSE in the fixed evaluation by 0.75% for RF, 2.59% for XGBoost, and 45.21% for LSTM. Across the walk-forward windows, it produces lower RMSE in seven of ten periods for RF and XGBoost and in nine of ten periods for LSTM. HAC-adjusted predictive-loss tests show statistically significant improvements for XGBoost and LSTM, while the RF differences are not statistically significant at the reference bandwidths. A time-ordered selection check further selects market_cap alone, indicating that the strongest reduction effect is concentrated in a highly parsimonious specification. With a publication-date and vintage-aligned information set, eight predictors (instead of nine) are selected and predictor reduction again lowers aggregate forecast errors for all three models. Predictor reduction therefore improves forecast accuracy in most comparisons within the tested learning models, although the random-walk benchmark produces lower absolute forecast errors than all learning-model specifications.