Hybrid ARMA-GARCH and Neural Networks for Market Shock Forecasting

Summary

The document describes a hybrid approach that combines ARMA-GARCH models with several neural network types to identify market patterns and predict the direction of shocks in defense equities and foreign exchange. ARMA-GARCH models capture linear time-series behavior and changing volatility; feature selection methods rank relevant inputs, while rolling training and validation are used to choose neural network forecasts. The compared networks include standard, recurrent-convolutional, decision-tree, and quantum variants.

The reported analysis uses daily defense-index data from 2022 and pound and yuan exchange-rate series from 1990–2016. It says conditional mutual information feature selection generally outperformed minimum-redundancy maximum-relevance, and threshold voting improved directional accuracy for defense equities; the reported averages were modestly above 50%. The quantum recurrent variant is described as especially strong. These are directional prediction results on specified datasets, not evidence of profitable trading after costs. The document provides limited detail on validation design and does not establish robustness across other periods or markets.

Key ideas

• ARMA-GARCH combines linear time-series modeling with conditional volatility estimation.
• The study pairs that baseline with several neural network architectures to predict market shock direction.
• Feature selection uses relevance and redundancy methods, with conditional mutual information reported as stronger.
• Rolling model selection and threshold voting are used to refine forecasts.
• Reported defense-market directional accuracy is only modestly above 50%, and profitability after trading costs is not shown.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.