Data Preparation

Log returns are acknowledged as mathematically superior, but percentage change is chosen to preserve the basic-arithmetic constraint. StandardScaler should be fit only on the training data and then applied to the test data to prevent leakage. Percentage returns are used because raw dollar price changes are not comparable across the full Bitcoin time series. Class balance must be checked before training because severe imbalance can make a majority-class model appear accurate without predictive power. Timestamp columns must be converted to datetime values to extract calendar features such as weekday. Class balance should also be checked after the train/test split to ensure the split did not create imbalance. Financial time series data must be split in temporal order and not shuffled. The analysis uses a Bitcoin/USD 12-hour OHLCV time series from 2023 to early 2026. Only the close price is used to construct the trading signal. The out-of-sample test set is intended to simulate deployment because the model never sees that period during training.