Intraday Volatility Regimes via Unsupervised Machine Learning: Evidence from the NIFTY Index

We propose an unsupervised machine learning framework for identifying structuredintraday volatility regimes in high-frequency financial data. Using 5-minute observations of the NIFTY index over the 2025 calendar year, we construct price-basedfeatures capturing realized volatility, volatility dynamics, and intrabar geometry.Nonlinear dimensionality reduction via Uniform Manifold Approximation and Projection (UMAP) is combined with density-based clustering using HDBSCAN to detect latent volatility states without imposing parametric distributional assumptions.The empirical analysis reveals a highly persistent baseline volatility regime punctuated by short-lived elevated volatility states. Estimated transition probabilities exhibit strong consistency with first-order Markov dynamics, and theoretical expectedregime durations align closely with observed persistence patterns. Volatility-relatedfeatures differ significantly across regimes, while return differences remain statistically insignificant, indicating that regime organization is primarily volatility-drivenrather than drift-driven.Out-of-sample evaluation demonstrates that regime-conditioned forecasts reduce30-minute realized volatility mean squared error by approximately 5.65% relative toa naive unconditional benchmark. These findings indicate that nonlinear unsupervised learning methods can uncover economically meaningful volatility structure andprovide incremental short-horizon predictive value in high-frequency equity markets.