Skip to content

Towards Robust Predictions Under Extreme Markets: A Hybrid Framework Integrating Adaptive Trend Decomposition and STOA‐Optimized Deep Learning

Sep 2026 · Expert systems · Vol 43 · 0 citations · 37 references

TL;DR

The proposed AT‐STOA‐Transformer‐BiLSTM hybrid framework offers improvements in predictive accuracy, stability and applicability, offering a promising tool for quantitative finance and risk management.

Abstract

Financial time series prediction faces dual challenges: deep learning models are highly sensitive to hyperparameters, and their vulnerability under extreme market conditions leads to severe performance decay. To address this, this paper proposes the AT‐STOA‐Transformer‐BiLSTM hybrid framework. The framework employs the STOA (Sooty tern optimization algorithm) for two‐stage collaborative optimization: first selecting predictive features, then tuning six key hyperparameters of the Transformer‐BiLSTM architecture. Concurrently, it integrates an AIC‐based adaptive trend decomposition module to separate non‐stationary sequences into deterministic trends and stationary residuals. In systematic experiments on China's CSI 300, CSI 500 and CSI 1000 indices, this framework demonstrates significant superiority. Compared to the unoptimized baseline, the entire model reduces the average absolute error (MAE) by approximately 24% on average and increases the coefficient of determination (R2) by 0.043 on average. In most cases, STOA is always superior to or on par with the mainstream meta‐heuristic optimizers (COA, DBO, EO, SO and SSA). Crucially, the framework exhibits considerable robustness. During the 2022 Russia–Ukraine conflict—a proxy for extreme market stress—the model maintains stable predictions, significantly outperforming baselines in the decay ratio metric. Furthermore, tests on the Dutch AEX and Swiss SMI indices confirm strong cross‐market generalization capabilities. In conclusion, the proposed hybrid framework offers improvements in predictive accuracy, stability and applicability, offering a promising tool for quantitative finance and risk management.

View source

Similar papers

Review Open access Sep 2026

Method evolution and comparison in asset price prediction: a review from traditional statistical models to deep learning

Asset price prediction is one of the most challenging research topics in finance. The inherent nonlinearity, low signal-to-noise ratio, and non-stationarity of Financial Time Series (FTS) make it difficult for traditional time series models to perform adequately. This paper reviews the evolution from traditional statis...

Yu-Nie Ng · 0 citations
Open access 2026

CSSFormer-VaR: A Cross-Scale State-Space Deep Learning Framework for Robust Financial Value-at-Risk Forecasting

The model introduces a Mamba-Enhanced Multi-Scale Bidirectional Fusion module (MEMBA-F) and a Cross-Scale Contextual Attention Gating module (CSCAG) to jointly model global temporal dependencies and local fluctuation features, thereby enhancing the identification capability for extreme risk events.

Wei-Yan Tan · 0 citations
#artificial intelligence Open access Jul 2026

A Hybrid LSTM–XGBoost Framework for Multi-Horizon Stock Return Prediction Across Diversified Equity Portfolios

Accurate prediction of equity returns remains a major challenge in computational finance due to the non-stationary, nonlinear, and low signal-to-noise ratio nature of financial time series. This paper proposes a hybrid two-stage architecture that combines a long short-term memory (LSTM) network with an XGBoost gradient...

Seifeldin Mostafa, Yahia Ahmed, Farah Darwish et al. · 0 citations
Open access Oct 2026

Residual Deep Learning for Realized Volatility Using a Hybrid HAR-LSTM Model with ARIMA Features and SHAP Interpretability

Classical linear models such as the Heterogeneous Autoregressive Realized Volatility (HAR-RV) model are difficult to beat at short horizons, yet cannot capture nonlinear, regime-dependent dynamics, while pure deep learning approaches typically overfit. This paper proposes a residual-learning hybrid in which a Long Shor...

C. Spulbar, Cezar Cătălin Ene · 0 citations
Open access Oct 2026

Meta-Learning Framework for Architecture and Data-Driven Loss Function Selection in Univariate Time Series Forecasting

The selection of loss functions in deep time series forecasting often defaults to standard metrics such as MSE or MAE. Previous studies have largely focused on average predictive accuracy using classic architectures, ignoring the interactions between optimization objectives, architectural inductive biases, and temporal...

Kana Kurniawan, E. Ripanti, Redi Ratiandi Yacoub · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.