BITCOIN PRICE PREDICTION WITH TECHNICAL INDICATORS: A HYBRID TRANSFORMER-RIDGE REGRESSION APPROACH

Abstract

Bitcoin has emerged as the dominant cryptocurrency, exhibiting rapid adoption alongside extreme price volatility that complicates investment strategies, risk management, and regulatory decision-making. While prior hybrid studies have predominantly combined multiple deep learning components such as CNN–LSTM or Transformer–GRU architectures, the integration of a deep neural architecture with a regularized linear model remains underexplored in Bitcoin price forecasting. To address this gap, this study proposes a hybrid framework combining a Transformer neural network with Ridge Regression, wherein the Transformer captures nonlinear temporal dependencies while Ridge Regression introduces L2 regularization to mitigate overfitting and enhance interpretability—an integration explicitly motivated by the bias–variance trade-off. The model is trained on technical indicators including MACD, Bollinger Bands, and RSI, and an ensemble weighting parameter α is systematically optimized via grid search. Empirical evaluation demonstrates that the hybrid model consistently outperforms standalone baselines, achieving an MAE of 1,251.572, RMSE of 1,623.004, R² of 0.991, and MAPE of 1.701%, with performance differences confirmed statistically via the Diebold–Mariano test. Economic validation reveals that the hybrid model is the only strategy to demonstrate statistically significant directional accuracy, although absolute trading returns remain below passive benchmarks under trending market conditions—a dissociation consistent with established findings in financial forecasting research. These results indicate that the model's primary contribution lies in forecast reliability and directional signal quality rather than return maximization under simple trading rules. Sensitivity to macroeconomic shocks and computational demands remain limitations for real-time deployment, suggesting directions for future research.