In the evolving landscape of edge devices, unlocking artificial intelligence (AI) services via the cloud has consistently raised concerns related to energy consumption, latency, and privacy. Shifting these services locally onto edge devices powered by CPUs and GPUs have always been hindered by their limited storage and computational power. This intensifies the demand for custom-designed AI accelerators. In this work, we propose an FPGA-based AI accelerator capable of efficiently processing time-series information. The proposed accelerator is built around a core algorithm modeled by a new variant of the standard gated-recurrent unit (GRU), namely random gated-recurrent unit (RGRU), thereby offering low latency, minimized resource utilization, and high performance compared to existing GRU accelerators. These improvements are achieved through parameter quantization, reduced computational overhead, and data localization, along with a 49 reduction in memory footprint compared to the standard GRU. The proposed design is evaluated on forecasting tasks using several standard benchmarks, including Yahoo Finance, Electricity Load, and Mackey-Glass, and demonstrates performance comparable to the standard GRU while requiring significantly fewer hardware resources.
Stock price time-series presents inherent difficulties due to their nearly random-walk behavior characterized by low signal-to-noise ratios and frequent regime shifts, whereas distinct industries add further complexity through unique volatility profiles and driving factors, which render cross-sector generalization especially demanding for predictive models. This study introduces IntraFormer, a novel Transformer architecture designed specifically for multivariate stock price prediction, integrating specialized intra-series attention mechanisms that enforce strict temporal causality, prioritize short-range dependencies, and incorporate gated residuals to combat noise overfitting while enhancing robustness against non-stationarity. Comprehensive experiments on 30 stocks spanning technology, healthcare, and financial sectors demonstrate IntraFormer's clear dominance: it consistently delivers the lowest MAE and RMSE, reducing them by 70%-90% compared with state-of-the-art Transformer variants, whose MAPE often exceeds 60%-99% in volatile cases; relative to recurrent baselines, IntraFormer achieves 20%-60% lower errors overall, with 30%-50% MAPE improvements. IntraFormer's portoflio construction returns nearly 11% on investment. This work establishes IntraFormer as a generalizable framework, offering potential for improving accuracy in trading strategies and portoflio risk managment across diverse market conditions.
This article innovatively proposes a stock return direction forecasting method by utilizing good and bad information spillover networks. First, we construct multilayer networks based on CHOL log return decomposition and Elastic-Net-VAR methods and analyze differences and similarities between bad and good information networks. It is found that the bad and good information networks exhibit various topology properties and evolution behaviors. Specifically, network connectedness of bad and good information layers changes over time, and during periods of financial stress, the interconnectedness between stocks increases dramatically. Besides, compared with other models, the empirical results show that the random forest (RF) prediction model outperforms other methods, and the forecasting power of the RF model is not only statistically significant but also economically valuable. Finally, a range of methods are applied to enhance the interpretability of the RF model, and we found that the core network characteristics that influence stock direction forecasting.
ABSTRACT Credit scoring requires modeling complex nonlinear relationships between borrower characteristics and default risk, yet highly expressive predictive models often suffer from limited transparency and weak generalization on small‐scale financial datasets. To address this challenge, this study proposes an interpretable Dual‐Rule Logical Neural Network (DLNet) for credit risk assessment. DLNet first converts continuous features into explicit binary threshold conditions, and then learns risk patterns through differentiable logical operators that explicitly construct conjunctive and disjunctive rule structures. To further improve rule compactness and readability, an explicit logic‐level sparse regularization term is imposed on the continuous logical weights during training, encouraging concise logical connections without penalizing the final prediction layer. In addition, a local rule backtracking extractor is developed to generate instance‐level explanations by tracing the top‐contributing activated logical nodes back to their corresponding threshold or categorical conditions. Experiments on four real‐world credit datasets show that DLNet achieves competitive predictive performance while providing both global rule patterns and local sample‐specific rationales. The learned rules further reveal heterogeneous logical structures across datasets, with disjunctive, conjunctive, and atomic patterns all appearing among high‐contribution rules. These results demonstrate that DLNet offers a practical framework for combining predictive effectiveness, structural sparsity, and decision transparency in credit scoring.
Accurate volatility forecasting in the European Union Emissions Trading System is crucial for risk management and policy design. However, conventional jump detection methods are prone to false discoveries, which can severely undermine predictive performance. To address this, we apply a universal threshold approach to reliably distinguish statistically significant true jumps from noise-induced spurious jumps, and develop a suite of extended HAR and LHAR models that incorporate these disentangled volatility components. Using high-frequency EUA futures data, our empirical analysis shows that models incorporating only true jumps consistently achieve superior in-sample fit and out-of-sample forecast accuracy. In contrast, including false jumps yields no meaningful improvement in predictive performance. These findings underscore that robust jump identification is essential for enhancing volatility forecasts. True jumps are predominantly associated with policy announcements and energy market shocks, whereas false jumps largely reflect market microstructure noise. This study offers empirical insights for carbon market volatility modeling and delivers actionable guidance for risk management and trading strategies.
