AFD-Net (Attention-Enhanced FAN with DLinear) is introduced, a novel deep learning framework designed to synergistically integrate frequency-domain decomposition, attention mechanisms, and linear time series modeling that consistently outperforms baseline models across multiple forecasting horizons.
Abstract
Exchange rate forecasting remains a fundamental challenge in financial time series analysis due to the inherent non-linearity, non-stationarity, and complex temporal dynamics of currency markets. This paper introduces AFD-Net (Attention-Enhanced FAN with DLinear), a novel deep learning framework designed to synergistically integrate frequency-domain decomposition, attention mechanisms, and linear time series modeling. The key architectural innovation lies in the sequential pipeline design: the FAN module first transforms non-stationary inputs into stationary residuals, enabling the subsequent attention module to operate in a low-noise environment where temporal dependencies are more reliably captured, and the DLinear backbone then decomposes the attention-enhanced representation into trend and seasonal components for final prediction. This causally motivated design ensures that each component addresses a distinct and well-defined challenge in financial time series modeling. Specifically, the model employs Feature-wise Adaptive Normalization (FAN) to isolate dominant frequency components, thereby mitigating non-stationary patterns effectively. Concurrently, an attention module is incorporated to adaptively emphasize salient temporal features, while a DLinear backbone serves as the core predictive structure, ensuring efficient sequence modeling with minimal computational overhead. We evaluate the effectiveness of AFD-Net through comprehensive experiments on the Exchange-Rate and IMF Exchange Rate datasets, benchmarking against seven state-of-the-art models, including Informer, Autoformer, PatchTST, and iTransformer. The results demonstrate that AFD-Net consistently outperforms baseline models across multiple forecasting horizons, exhibiting particular superiority in long-term trend modeling. Furthermore, ablation studies quantitatively confirm the distinct and complementary contributions of the FAN and attention components. A comparison against ReVIN-based normalization further justifies the FFT-based design of FAN, demonstrating that the performance gain substantially outweighs its modest computational overhead. This work highlights the potential of frequency-aware architectures in enhancing the accuracy and robustness of financial time series predictions.
The extreme non-stationarity, high noise levels, and multi-timescale coupling in financial futures markets pose major challenges for time series forecasting. Existing models often struggle to disentangle localized shocks from global trends due to incompatible inductive biases. To address this issue, we propose a Decomposition-Enhanced Network (DENet). Following a divide-and-conquer paradigm, DENet adopts a multi-stream architecture: the main path extracts stable trends via moving averages and dual-path linear projections, while Auxiliary Stream I captures seasonal and local cyclical patterns using depthwise separable convolutions, and Auxiliary Stream II models high-frequency dynamics through a nonlinear autoregressive-style mapping. These components are integrated via an adaptive fusion mechanism, balancing global robustness and local structural sensitivity. Experiments on real-world futures data demonstrate that DENet outperforms a wide range of state-of-the-art benchmarks. Compared with seven state-of-the-art baselines, DENet achieves an average reduction of 10.15% in RMSE for daily forecasting on iron ore futures, and an average reduction of 22.57% in MAE under the 5-min, 12-step forecasting horizon. Furthermore, we integrate DENet’s dual-scale predictions into the R-Breaker strategy with parameter switching and dynamic position sizing. Backtesting results reveal that the annualized return of iron ore futures outperforms the baseline strategy by an average of 7.6 percentage points across five backtesting windows. Ultimately, DENet effectively bridges advanced structural modeling and actionable algorithmic trading.
Stock price forecasting remains challenging because financial time series usually exhibit nonstationarity, strong noise, nonlinear dynamics, multi-scale fluctuations, and abrupt short-term shocks. Although CEEMDAN-LSTM-Transformer-based decomposition–prediction–reconstruction frameworks can reduce the complexity of raw closing price sequences, their IMF branches usually perform one-shot open-loop prediction, where the historical prediction errors of each IMF branch are not explicitly exploited before final reconstruction. To address this limitation, this paper proposes FI-CAECNet, a frequency-aware IMF-wise closed-loop adaptive error correction network for stock price forecasting. In the proposed framework, CEEMDAN first decomposes the closing price series into multiple intrinsic mode functions and one residual component. Each IMF component, together with auxiliary financial indicators, is predicted by an LSTM branch to obtain an initial component-level prediction, while the residual component is modeled by a Transformer branch. The core innovation lies in the proposed FI-CAEC module, which is embedded after each IMF-LSTM branch. FI-CAEC uses historical error memory and frequency-aware features, including energy, entropy, volatility, and dominant frequency, to construct a feedback state. An adaptive correction gate and a bounded correction term are then generated to iteratively refine the initial IMF prediction in a closed-loop manner before final summation. Experiments on eight representative Chinese A-share and U.S. stock datasets show that FI-CAECNet achieves the best overall forecasting performance among all compared models. Compared with the CEEMDAN-LSTM-Transformer baseline, FI-CAECNet reduces the average MSE and MAE and improves the hit ratio. Dedicated ablation experiments further show that historical error memory, frequency-aware features, gated correction, bounded correction, and finite closed-loop iteration all contribute to the performance improvement. The results demonstrate that FI-CAECNet can effectively reduce component-level prediction errors and improve the accuracy, robustness, and interpretability of stock price forecasting.
Forecasting stock prices is challenging due to the non-stationarity and volatility of financial time series. We propose the Generative Adaptive Decomposition Hierarchical Transformer (GADHT), a hybrid framework that combines adaptive decomposition, masked self-supervised pretraining, and hierarchical attention. GADHT applies Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose financial signals into intrinsic mode functions (IMFs) that capture multi-scale temporal dynamics. A masked self-supervised pretraining task based on IMF reconstruction is used to learn spectral–temporal representations without labeled data, while a hierarchical transformer with energy-weighted attention emphasizes informative IMFs during forecasting. Experiments on large-cap equities across multiple forecasting horizons show that GADHT achieves competitive and stable forecasting performance. The model maintains stable predictive behavior during stress periods such as the 2008 financial crisis and the 2020 COVID-19 crash, and shows positive economic performance under the adopted backtesting assumptions. Zero-shot and cross-market experiments further suggest that the learned representations can transfer to unseen large-cap equities and selected international markets. Interpretability analysis shows that spectral attention patterns are associated with financial characteristics such as momentum, volatility, and liquidity. Overall, the results suggest that GADHT provides a coherent and interpretable framework for multi-horizon stock forecasting.
Khalid Bentaleb, Mohamed Ben Houad, Mohammed Mestari· IEEE Access· 0 citations
The Decomposed Recurrent Neural Network (DeRNN) is proposed, which decouples global trend modeling from local fluctuation extraction via an asymmetric dual-track architecture and exhibits superior robustness against noise and distribution shifts.
Shanyun Qian· Poster Volume 0008 The 2026...· 0 citations
Financial time series forecasting tasks, specifically stock trend prediction and market attribution, are essential for quantitative investment and risk control. However, these tasks suffer from low signal-to-noise ratios and non-stationarity, stemming from the coupling of hierarchical drivers: market trends, sector rotations, and idiosyncratic dynamics. Existing methods, often relying on static sector labels or time-domain correlations, struggle to capture dynamic, multi-scale dependencies and lack the interpretability required for return attribution. To address this, we propose ResDIF, a Residual Disentanglement framework for Interpretable financial time series Forecasting inspired by Asset Pricing Theory (APT). ResDIF employs a progressive residual architecture via a Spectrally-Enhanced Temporal Encoding mechanism to explicitly decompose stock data into market, sector, and individual layers. Furthermore, a self-supervised orthogonal loss encourages feature separation, enabling the model to autonomously decouple systematic risks from idiosyncratic alpha. Beyond predictive modeling, ResDIF effectively unifies high predictive accuracy with granular structural attribution. By leveraging this intrinsic interpretability to quantify structural market risks, our framework further enables adaptive portfolio optimization through dynamic hedging and asset selection. Experiments demonstrate that ResDIF outperforms existing methods while providing actionable interpretability support for refined portfolio risk management.
Chengwei Fu, Gang Xiao, Yuchao Zhang et al.· Proceedings of the 32nd ACM...· 0 citations
Misinformation events introduce abrupt, nonlinear distortions in financial markets, posing significant challenges to conventional time-series models that often fail to capture such structural breaks due to assumptions of stationarity and limited contextual awareness. To address this limitation, we develop an Enhanced Event-Adaptive Transformer (EET) architecture that jointly models global contextual dependencies and localized temporal dynamics under event-driven scenarios. The proposed framework integrates event embedding layers and multi-scale con volutional feature extractors into a Transformer encoder, enabling explicit encoding of exogenous disruption signals alongside multi-resolution pattern recognition. We evaluate the model using stock-level data from China's power sector (2017-2022), with the misinformation event of October 16th, 2019, serving as a causal boundary for temporal segmentation. Trained exclusively on pre-event data and tested on post-event sequences, the model achieves superior predictive performance in three-fold cross-validation (mean $\mathrm{R}^{\mathrm{2}}=\text{0. 9 8 8 4} \pm \text{0. 0 0 0 5}$; MAE ${=}\text{0. 1 1 7 2} \pm 0.0264)$. Ablation studies confirm that both event embedding and convolutional components significantly enhance accuracy (R2 declines by 21.0% and 17.1% upon their removal, respectively). Residual diagnostics validate statistical robustness, with errors exhibiting near-normal distributions and no temporal bias. This study establishes a novel theoretical and practical framework for quantifying the impact of exogenous shocks on financial systems. The architecture provides an interpretable and high-fidelity solution for event-driven forecasting, with implications for financial risk management and algorithmic decision-making.
Xuemin Deng, D. Wu· Annual International Compute...· 0 citations