From Trends to Transitions: ARIMA Powered by Hidden Markov Regimes for Adaptive Forecasting

Badrikanath Praharaj

Abstract:

This study presents an Adaptive ARIMA-HMM
framework for regime-aware forecasting, applied to the Indian
Sensex index. The model combines ARIMA’s linear time-
series prediction with probabilistic regime identification by
applying Hidden Markov Models to ARIMA residuals, enabling
the detection of latent market states. Using smoothed regime
probabilities, it dynamically adjusts equity exposure (100%,
50%, or 0%), allowing timely responses to market shifts. In
backtests, the model achieved a total return of 3040.88%,
annualized return of 40.44%, Sharpe ratio of 4.63, and
maximum drawdown of -19.62%, outperforming standalone
ARIMA (17.12%), ARIMA-LSTM (134.09%), and static hybrid
baselines (-26.65%). Sensitivity analysis confirms that a three-
regime structure offers optimal balance between risk and
return. While the framework improves interpretability and
regime adaptation over deep learning models, its reliance on full
historical data and absence of transaction cost modeling pose
real-world challenges. Nonetheless, the Adaptive ARIMA-HMM
offers a robust alternative to traditional and neural approaches,
particularly in volatile, data-limited emerging markets, where
macroeconomic regime triggers are sparse or noisy.