Regime-Aware Hierarchical Modeling for Intra-day VIX Dynamics

Forecasting short-horizon movements in implied volatility constitutes a persistent challenge in quantitative finance due to non-stationarity, volatility clustering, structural regime changes, and non-linear cross-market dependencies. This study proposes a novel regime-sensitive multi-model classification framework for the categorical prediction of intra-day directional movements in the CBOE Volatility Index (VIX). The architecture combines kernel-based discriminative models with non-parametric tree-based learners under a dynamic decision aggregation scheme specifically designed to adapt to heterogeneous volatility regimes. To operationalize the forecasting framework in a production-grade, real-time environment, this study implements an event-driven cloud-native data infrastructure based on Microsoft Azure Functions. A timer-triggered serverless orchestration layer manages the automated ingestion of multi-asset market data, the computation of derived volatility, term-structure, and cross-market features, and the persistent storage of processed observations within a cloud-hosted MongoDB NoSQL environment. The infrastructure is designed to support low-latency processing of high-dimensional financial time series while ensuring horizontal scalability, fault tolerance, and schema flexibility required by continuously evolving market datasets. Feature generation, validation, transformation, and persistence are fully automated within a modular ETL work-flow, enabling reproducible data acquisition and seamless integration with downstream inference pipelines. As a result, the proposed architecture provides an end-to-end deployment framework for continuous volatility forecasting under live market conditions. To address class imbalance and enhance predictive robustness, targets are defined using quantile-based categorical thresholds on intra-day VIX returns, ensuring balanced class representation across heterogeneous market conditions. Model hyperparameters are optimized within a walk-forward validation framework, preserving temporal dependencies while mitigating overfitting and preventing information leakage during calibration. Performance is subsequently assessed through an out-of-sample backtesting procedure spanning multiple test windows with varying horizons, enabling the evaluation of model stability, robustness, and generalization across structurally distinct market regimes. Empirical results indicate that performance gains are particularly pronounced during high-volatility regimes, highlighting the importance of explicitly incorporating regime sensitivity into volatility forecasting models.Across all evaluated forecasting horizons, the best-performing specification is a logistic regression model, which outperforms more complex non-linear learners, including tree-based ensembles, in out-of-sample backtesting. In addition, it exhibits strong stability across evaluation windows, achieving a median (Q2) weighted F1-score of 0.54 and a substantially lower variability with a standard deviation of 0.081. Overall, the empirical evidence suggests that the proposed architecture constitutes a statistically robust and com-putationally scalable framework for short-horizon volatility classification under non-stationary market dynamics. By integrating regime-sensitive machine learning with an event-driven cloud-native data infrastructure, the system achieves consistent out-of-sample generalization across heterogeneous backtesting horizons and structurally distinct volatility regimes. From a systems perspective, the serverless orchestration layer and distributed NoSQL storage architecture provide low-latency feature processing, fault-tolerant execution, and schema adaptability, while enabling seamless horizontal scaling as data dimensionality, asset coverage, and inference frequency increase. These properties establish a reproducible and production-ready computational framework for real-time volatility modeling, systematic execution, and adaptive risk calibration in high-frequency quantitative environments.