Machine Learning-Based Analysis of Stock Market Volatility and Investment Risk
DOI:
https://doi.org/10.71086/IAJE/V13I2/IAJE1310Keywords:
Stock Market Volatility, Investment Risk, Deep Learning, Regime Switching, Volatility Targeting, Drawdown Control.Abstract
Equity volatility clusters have a tendency to go from calm volatility to turbulent to crisis, and static risk models do not fare well with fat-tailed shocks. In this paper, a deep-learning architecture that predicts volatility and categorizes the market regime in one network is studied, and its results are used to facilitate investment risk control. The network is fed a 22-day window of return-based features, which are passed through an attention layer, a bidirectional LSTM, and a convolutional encoder, then split into a 5-day volatility head and a three-state regime head. Its outputs are used to control an allocation rule, which sets the size of exposure to a volatility objective, and then cuts it back when a crisis is likely to occur. There are 5 independent markets that are simulated and known to the evaluator. When compared with the baseline models, EWMA, GARCH, HAR, and random forest, the model reduces volatility forecast error by 5.1% to 16.6%, identifies the crisis state with an F1 score of 82.9%, and decreases the average maximum drawdown of the buy-and-hold strategy by 50.2% to 16.7%, while increasing the Sharpe ratio from 0.03 to 0.66. The difference between a random forest and attention is negligible and lies within one standard deviation. The evidence is represented in a simulated fashion, and the scope of the evidence is explicitly described.
