A Risk-Aware Hybrid Ensemble Approach for AEX Index Forecasting: Integrating APARCH-t Volatility with LSTM-CNN-RF Architectures

The accurate prediction of financial time-series remains a formidable challenge due to inherent non-linearity,
heteroskedasticity, and the presence of "fat-tailed" distributions. This study proposes a novel, risk-augmented hybrid
ensemble framework designed to enhance the forecasting precision of the Amsterdam Exchange (AEX) index.
Departing from conventional monolithic models, the research methodology integrates an asymmetric power
autoregressive conditional heteroskedasticity (APARCH) model with a Student-t distribution to extract robust volatility
features. These econometric inputs are subsequently fed into a tripartite deep learning ensemble comprising Long
Short-Term Memory (LSTM) networks for temporal dependencies, Convolutional Neural Networks (CNN) for spatial
feature extraction, and Random Forest (RF) for non-linear regression refinement.
Empirical results demonstrate that the proposed architecture significantly outperforms baseline models, achieving a
high predictive accuracy characterized by an $R^{2}$ of 0.9408 and a Mean Absolute Error (MAE) of 4.9542. A
critical finding of this research is the significance of the leptokurtic nature of AEX returns (Kurtosis: 12.64); by
anchoring the machine learning engine with APARCH-derived conditional volatility, the model effectively mitigates the
impact of market noise and transient shocks. Furthermore, Value-at-Risk (VaR) backtesting validates the model’s