Dynamic learning rate adjustment using volatility in LSTM models for KLCI forecasting

The prediction of financial market behaviour constitutes a multifaceted challenge, attributable to the underlying volatility and non-linear characteristics inherent within market data.  Long Short-Term Memory (LSTM) models have demonstrated efficacy in capturing these complexities.  This study proposes a novel approach to enhance LSTM model performance by modulating the learning rate adaptively based on market volatility.  We apply this method to forecast the Kuala Lumpur Composite Index (KLCI), leveraging volatility as a key input to adapt the learning rate during training.  By integrating volatility into the learning process, the model can better accommodate market fluctuations, potentially leading to more accurate and robust predictions.  The proposed dynamic learning rate adjustment mechanism operates by scaling the learning rate according to the most recent volatility measurements, ensuring that the model adapts swiftly to changing market conditions.  This approach contrasts with traditional static learning rates, that may fail to sufficiently account for the dynamic of financial markets.  We conduct extensive experiments using historical KLCI data, comparing our proposed model with standard LSTM and other baseline models.  The results demonstrate that our volatility-adjusted learning rates outperform conventional LSTM models with fixed learning rates with respect to predictive performance and stability.  The findings suggest that incorporating volatility into learning rate adjustments can significantly enhance the predictive capability of LSTM models for stock market forecasting.  The improved forecasting accuracy of the KLCI index highlights the potential of this approach for broader applications in financial markets.