QUANT REGIME 50: A PREDICTIVE MACHINE LEARNING FRAMEWORK FOR STRUCTURAL REGIME-SWITCHING AND MARKET RISK ANALYTICS ON THE NIFTY 50 (NSEI)

Abstract

Financial time-series data—particularly stock market index returns across emerging markets—frequently exhibit non-stationarity, time-varying volatility clustering, and sudden structural shifts. Traditional investment and risk management models rely heavily on static covariance assumptions, often leading to severe capital drawdowns during market downturns. To address these structural flaws, this project introduces QuantRegime-50, a quantitative framework that integrates unsupervised econometrics, supervised machine learning, and dynamic tail-risk analytics tailored for the Nifty 50 Index (NSEI). The framework operates via a two-stage hybrid pipeline: 1. Regime Identification: An unsupervised Hidden Markov Model (HMM) isolates hidden market states directly from daily log returns. The model identifies two distinct environments: a low-volatility expansion phase (Regime 0: 𝜇0 = +0.0058%/day, � �0 = 0.647%) and a high-volatility contraction phase (Regime 1: 𝜇1 = −0.0218%/day, 𝜎1 = 1.552%). 2. Predictive Classification: These extracted states act as target labels for a Random Forest classifier trained on rolling technical indicators, including RSI, MACD, volatility, and Bollinger Bands. The model achieves an out-of-sample classification accuracy of 74% in forecasting next-day regime transitions. Out-of-sample back testing shows that the regime-aware trading strategy outpaced the static Buy & Hold benchmark, delivering a cumulative return of 1.77% compared to 1.12% for the index while significantly cutting downside risk drag (improving the Sortino Ratio from -0.85 to -0.37). Additionally, dynamic Value at Risk (VaR) and Conditional Value at Risk (CVaR) calculations adapt immediately during state changes, providing an automated layer of defence to protect capital before severe drawdowns take effect.