Volatility-adaptive temporal learning framework using hybrid ARIMA-QuadGRU attention for multi-regime financial forecasting

Abstract

Conventional financial market forecasting models are challenged by the non-stationarity, the existence of regime changes, the presence of structural breaks, and the phenomena of volatility clustering in financial markets. Linear approaches have only limited accuracy in terms of direction, as they do not capture the nonlinear behavior and long-term dependence. Given the above challenges, we introduce a multi-regime aware deep temporal forecasting framework, which combines the statistical modeling technique autoregressive integrated moving average (ARIMA) with a novel quad-gated recurrent unit (QuadGRU) and multi-head attention (8 heads) in a hybrid way, depending on the volatility of the data. The framework is tested using 15 years of daily market price data from the NIFTY-50 index from Yahoo Finance API (2010–2025). Our pipeline systematically preprocesses the data by normalizing the data and masking filtering anomalies using the isolation forest (IsFo) algorithm with training-only fitting, and finally, we conduct regime-dependent feature engineering using technical indicators such as moving average, Relative Strength Index, and momentum oscillators. Gradient boosting machines (GBM) can be used for feature importance ranking and principal component analysis (PCA) for orthogonal representation of the features. The Granger causality test is used to define the temporal causality between predictors before sequence modeling. ARIMA is used for short-term linear component and QuadGRU is used for short-term nonlinear component of temporal dependency, the hybrid forecasting architecture is used. The novel QuadGRU gating mechanism offers adaptive memory regulation over the entire range of volatility, and multi-head attention delivers interpretable adaptive weighting of the historical information. The rolling window validation experiments with the baseline models ARIMA, LSTM and GRU improvements in MSE (0.012 ± 0.003 in normalized [0,1] scale, corresponding to 14,850 index-point MSE), RMSE (0.110 ± 0.014 normalized, corresponding to 121.8 index points), MAE (0.084 ± 0.011 normalized, corresponding to 93.1 index points), and directional accuracy (87.5 ± 2.1%), especially for time periods with high volatility. Error variance distribution analysis, bootstrap forecasting error distributions, volatility-conditioned error analysis, Granger causality validation, anomaly detection rate and false positive rate, as well as prediction stability metrics have been used for statistical validation of our performance improvements. The results justify the proposed regime switching hybrid approach as a scalable forecasting methodology for financial time-series accounting for the regime change in market conditions and provide investors, portfolio managers, and financial analysts with a practical tool to support their investment decisions in volatile markets.