Evaluasi pendekatan sliding window terhadap static split dalam prediksi harga Bitcoin menggunakan algoritma Random Forest

Abstrak

Bitcoin price prediction remains a complex and relevant challenge due to the asset’s high volatility and increasing adoption across sectors. One important factor influencing prediction performance is the data-splitting strategy used during model development. This study compares two approaches (sliding window and static split) in the context of Bitcoin price forecasting using the Random Forest algorithm. Historical data from Yahoo Finance spanning 2015 to 2022 is used, with input features constructed from closing prices and seven daily lags. The sliding window method trains the model with a moving 365-day window and tests it on the following day, whereas the static split uses a fixed date-based partition. Evaluation is conducted using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), R-squared (R²), Mean Absolute Percentage Error (MAPE), and direction accuracy metrics. The results indicate that the sliding window approach produces more accurate and consistent predictions and better captures directional trends in price movement than the static split method.