Forecasting Volatility Using Hybrid Machine Learning Method: Sequencing Block, Multi-layer Perceptron, and Bayesian Optimization

Financial markets are characterized by their inherent complexity, nonlinearity, and stochastic behavior. Traditional econometric models, including generalized autoregressive conditional heteroskedasticity and autoregressive integrated moving average approaches, often struggle to capture the dynamic patterns present in financial time series. Similarly, conventional machine learning techniques such as decision trees, random forests, and support vector machines exhibit limitations in terms of generalization and adaptability to shifting market conditions. This study introduces a novel hybrid forecasting framework that integrates a sequencing mechanism, a deep feedforward neural network with multiple layers, and a Bayesian optimization strategy to enhance predictive performance and model robustness. The proposed approach employs Bayesian optimization to systematically fine-tune both the temporal feature extraction parameters and the neural network’s hyperparameters, ensuring an adaptive and comprehensive modeling configuration. When applied to forecasting on data from the S&P 500 index with a short-term prediction horizon, the hybrid model demonstrates marked improvements in predictive accuracy and model fit compared to traditional machine learning benchmarks. It also effectively mitigates overfitting and offers greater computational efficiency than more complex deep learning architectures, such as recurrent neural networks and generative adversarial models.