<![CDATA[The primary challenge in cryptocurrency price prediction lies in the market’s highly volatile, non-linear, and near–random walk behavior, which makes traditional predictive models unable to achieve consistent accuracy. This study aims to develop and evaluate a hybrid model combining Long Short-Term Memory (LSTM) and XGBoost to predict price direction and returns for ten major cryptocurrencies using daily data from 2023 to 2025. Historical data were processed through feature engineering, normalization, and sliding-window sequence construction, and the models were evaluated using TimeSeriesSplit to prevent data leakage. The results show that the hybrid model consistently outperformed both LSTM and XGBoost, achieving an average directional accuracy of 58.6%, significantly higher than the baselines (51.7% for LSTM and 53.6% for XGBoost). The average RMSE of 0.0289 indicates stable return predictions without systematic bias. Statistical validation through paired t-tests and McNemar tests confirmed the significance of the improvement at p < 0.001. A trading simulation using a 1-day holding period produced an annualized return of 41.5% with a Sharpe ratio of 1.12, outperforming the buy-and-hold strategy. These findings highlight that integrating LSTM’s temporal representation with XGBoost’s non-linear learning capabilities is an effective and computationally efficient approach for cryptocurrency price forecasting, offering practical value for the development of algorithmic trading systems.]]>
