When Does Model Complexity Pay Off? Boosting, Recurrent, and Hybrid Architectures for Volatility-Informed Portfolio Allocation in the Indian Equity Market

Abstract

To ensure dynamic portfolio allocation and real-time risk management, it is crucial to accurately forecast volatility; this is especially important in broadly distributed, emerging equity markets that display both structural inefficiencies and high levels of intraday volatility. Although practitioners are increasingly adopting complex deep learning architectures for financial forecasting, it remains unclear whether the added architectural complexity of sequential models such as Long Short-Term Memory (LSTM) networks — or of hybrid and ensemble constructions built on top of them — translates into materially better forecasting accuracy or portfolio outcomes than simpler, structured tree-based methods (e.g., Gradient Boosting) in high-frequency datasets. Therefore, this study conducts a systematic out-of-sample evaluation of forecasting accuracy against model complexity, using 15-minute intraday data for 20 large-cap Indian equities traded on the National Stock Exchange between 1 March 2019 and 12 October 2023. Subsequently, the respective models predicted volatility will be incorporated into risk-adjusted portfolio optimization strategies; specifically, Minimum Variance, Risk Parity, Mean-Variance, Adaptive Rebalancing, and Conditional Value at Risk (CVaR). The study employed a rolling, walk forward evaluation framework in order to assess the relative performance of the respective models and to identify whether the implementation of hybrid GB-LSTM and ensemble models will provide incremental predictive improvement in comparison to the respective models in aggregate with respect to predicted volatility. The main findings indicate that added model complexity does not translate into proportionate forecasting gains in this setting: the simpler, structured Gradient Boosting approach matches or exceeds the accuracy and stability of the more complex recurrent and hybrid architectures, while ensemble averaging recovers most of Gradient Boosting's accuracy without requiring a fully specified deep sequential model. The findings are applied to evaluate the marginal economic value of complexity and not intended to identify a winning architecture. Back-testing results of the portfolio indicate that by applying boosting models when predicting volatility, one may increase the risk-return ratios of the portfolios applying CVaR and minimum variance portfolio management. In particular, risk-return indicators generated during the testing experiments with the use of the boosting models turned out to be the best in terms of Sharpe and drawdown ratios under the constrained turnover and exposure levels. Therefore, this study contributes into discussions on model selection for financial time series forecasting by providing evidence from the emerging market in favor of the point that the complexity of the architecture does not necessarily correspond with economic efficiency of the generated financial forecasts. Additionally, the scope of application of findings is extended to quantitative asset managers, financial technology companies, and researchers working on the systems based on the assessment of asset allocation skills in the market rich in data but plagued by structural noise.