| Abstract [eng] |
This thesis examines whether regime awareness and alternative data improve next-day Bitcoin log return forecasting. A daily dataset for 2019–2024 is constructed by combining market data (OHLCV and technical indicators) with on-chain activity, sentiment measures, and macro-financial variables, while regimes are included as an additional conditioning signal. Models are evaluated using a strict chronological split with a held-out 2024 test set. The study compares a naive zero-return baseline against LSTM based sequence modeling, LightGBM regression on engineered predictors, and a hybrid approach that augments LightGBM with LSTM embeddings. Performance is assessed using RMSE, MAE, and directional accuracy. Results indicate that performance differences across model families are small and remain close to the naive benchmark, highlighting the limited predictability of daily Bitcoin returns. The best configuration provides only marginal RMSE improvements and directional accuracy remains near 0.5, suggesting limited standalone trading value but supporting the use of the framework for regime conditioned interpretation and risk oriented extensions. |
