Advanced Time Series Modeling for High-Risk Financial Instruments Using VRA-Enhanced Machine Learning

Financial markets face frequent volatility and rapid regime shifts, which lead to uncertainty and risk for investors and financial institutions. This can be addressed by building a model that remains reliable under changing conditions and detects rare but impactful high-risk events. However, building such a model constitutes a great challenge. Individual Machine Learning (ML) models can capture complex nonlinear patterns. However, they often face a trade-off between recall and precision as models optimized for recall tend to produce many false alarms, whereas models with higher accuracy may miss significant risk episodes. The current study addresses this issue by proposing the Volatility Risk Analyzer (VRA), a hybrid prediction framework combining both supervised and unsupervised learning techniques. Unsupervised regime detection is performed using K-means clustering, and Long Short-Term Memory (LSTM) networks are utilized to model temporal dynamics in financial time-series data. The clustering outputs are integrated employing a hybrid training technique to exploit their complementary strengths. The proposed framework is evaluated using an external dataset of five years of ICICI Bank OHLCV stock data, supplemented with widely deployed technical indicators such as Moving Averages, Moving Average Convergence Divergence (MACD), Rate of Change (ROC), and Williams %R. The results indicate that combining supervised and unsupervised learning enhances volatility risk prediction, offering a novel contribution for both academic research and real-world risk management applications.