Aug 2026· AI and Data Science Journal· 0 citations· 13 references
TL;DR
The findings indicate that passing attention-derived context into a bidirectional memory module offers a practical means of combining long-horizon structure with local temporal variation, although computational cost remains relevant for latency-sensitive trading applications.
Abstract
Forecasting financial indices remains difficult because market observations combine persistent movements, short-lived disturbances, nonlinear interactions, and substantial noise. A single recurrent architecture may retain nearby temporal information yet fail to represent widely separated dependencies, whereas an attention-based encoder is effective at relating distant positions but does not by itself guarantee detailed sensitivity to local sequence dynamics. This paper reformulates the forecasting pipeline as a cascaded global–local learning problem. A Transformer encoder first converts normalized price windows into contextual representations through positional encoding, multi-head self-attention, residual normalization, and a feedforward sublayer. Those representations are then processed by a bidirectional long short-term memory network, so forward and reverse recurrent states refine the global context before a dense prediction head produces the output. The design was assessed on daily closing-price series for the S&P 500, Dow Jones Industrial Average, and Nasdaq Composite over 2 September 2003–13 July 2023. Preprocessing included interpolation of missing observations, interquartile-range screening of outliers, first differencing, min–max scaling, and windowed sample construction. Tests against recurrent, bidirectional recurrent, feedforward, Informer, and temporal-convolution baselines show that the hybrid system delivers the strongest overall error and goodness-of-fit profile across the three markets. The findings indicate that passing attention-derived context into a bidirectional memory module offers a practical means of combining long-horizon structure with local temporal variation, although computational cost remains relevant for latency-sensitive trading applications.
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
The results indicate that self-supervised representation learning can serve as an effective and scalable substitute for manual feature engineering in finance time-series forecasting.
M. Pawar· International Journal of Ada...· 0 citations
Stock index forecasting is essential for financial market analysis and risk monitoring, yet it remains challenging because index price series are nonlinear, non-stationary, and driven by heterogeneous market factors. Existing methods remain limited in preserving local price patterns, capturing multi-scale local dependencies, and integrating attention-derived structures with long-range state-space representations. To address these limitations, we propose AG-SSM, an attention-guided state-space model for multi-step stock index forecasting. The model first uses variable-wise patch embedding to construct local semantic units, which are then processed by the AG-SSM architecture for temporal representation learning. Its core block integrates dual-path local attention (DPLA), S4D-based state-space feature generation, attention-guided aggregation (AGA), and gated update (GU). Specifically, DPLA combines sliding and dilated local attention to capture contiguous and sparsely distributed dependencies, while AGA reuses local attention maps to refine state-space features, thereby coupling local market structures with long-range sequential dynamics. Experiments on six stock index datasets (SSE, SZSE, SMESE, SP500, DJIA, and NIKKEI225) under one-, five-, ten-, and fifteen-step forecasting horizons show that AG-SSM achieves the lowest horizon-averaged MAPE on all six datasets while maintaining competitive performance across other metrics and individual horizons. Averaged over five independent runs, the horizon-averaged MAPE values are 1.5466%, 2.2173%, 2.2293%, 1.4373%, 1.2995%, and 1.8738% on the six datasets, respectively. Ablation studies, state-space variant comparisons, sensitivity analyses, and statistical tests further support the effectiveness and robustness of the proposed framework.
Hao-Rong Liao, Xiang-Zeng Kong, Yiming Mu et al.· Mathematics· 0 citations
Stock market price prediction remains a persistent challenge owing to the non-stationarity, high noise content, and intricate spatiotemporal dependencies that characterize financial time series. Existing approaches typically excel at either local pattern extraction or long-range dependency modeling, yet seldom reconcile both within a unified framework. This paper introduces a CNN–Transformer dual-channel architecture equipped with a dynamic attention fusion module for stock price forecasting. The convolutional channel applies hierarchical dilated convolutions to distill fine-grained local patterns from multi-indicator sequences while suppressing high-frequency noise. Simultaneously, the Transformer channel employs multi-head self-attention to capture long-distance temporal correlations and regime-shift dynamics. A learnable gating mechanism then fuses the two feature streams by adaptively weighting local detail against global trend information according to market conditions. Experiments conducted on four real-world stock datasets spanning the S&P 500, CSI 300, NASDAQ Composite, and Hang Seng Index show that the proposed model reduces mean absolute error by 9.7–15.3% and root mean square error by 9.5–13.8% relative to competitive baselines including LSTM, CNN–LSTM, Informer, and PatchTST. Ablation studies further indicate that both channels and the fusion module contribute to prediction accuracy, and the architecture remains effective across markets with differing volatility profiles.
This study proposes MSATE-Net for next-day stock index forecasting. The model combines parallel one-dimensional convolutions with receptive fields of 3, 7, and 15 trading days, a bidirectional LSTM operating entirely inside a historical lookback window, sample-dependent temporal attention, and a residual regularized prediction head. Here, “bidirectional” denotes paired processing of the same observed window; it does not assert time-reversal invariance of financial prices or access to observations after the forecast origin. The globally learned attention temperature controls overall selectivity and is not described as a regime-specific adaptive parameter. Experiments use S&P 500, CSI 300, and Nikkei 225 data; persistence and drift benchmarks; recent forecasting architectures; five-seed uncertainty estimates; expanding-window tests; return and directional metrics; and Diebold–Mariano comparisons. The revised evidence supports lower price-level errors, while directional and significance results are mixed across markets. Because a separate model is fitted in each market, the findings establish cross-market consistency rather than transfer learning.
Forecasting time series over long horizons is essential for proactive decision-making in many systems. Recent research has focused on transformer-based architectures, which capture long-range dependencies in sequential data. However, several studies show that simpler linear models can outperform transformers by avoiding overfitting during training. In this context, we present NeuroFlexMLP, a deep learning model for multivariate time series forecasting tasks. NeuroFlexMLP's key distinct feature is the adaptability to the diverse complexity of real-world time series, which is achieved, from the architecture standpoint, by adding non-linear residual blocks to a first linear block. This architectural design simplifies hyperparameter optimization, leading to accurate forecasts for various time series data types regardless of the lookback or prediction horizons, outperforming state-of-the-art (SOTA) models on challenging real-world datasets. Its Multi-Layer Perceptron (MLP) design ensures high computational efficiency, making it scalable for longer input sequences than transformer-based models. We validate NeuroFlexMLP for the LEO satellite beam hopping use case, where its lightweight design enables on-board deployment, and on state-of-the art AI datasets. Across all these benchmarks, NeuroFlexMLP achieves competitive accuracy over state-of-the-art models while providing an adaptive architecture that significantly reduces computational overhead. On the LEO beam hopping task, it achieves up to 35.9% MSE reduction over Informer, which translates into up to 28% lower provisioning cost under asymmetric cost models that penalize under-allocation more heavily than over-allocation.
P. F. Pérez, Claudio Fiandrino, Marco Fiore et al.· La Main· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
