Scholarlink

<p>Early identification and reliable early warning of financial risks are critical challenges for maintaining market stability and serving financial security strategies. Addressing the limitations of existing models—including insufficient accuracy, weak interpretability, and high deployment costs under high-noise, strong nonlinearity, and low-latency constraints—this paper proposes Light-FinTS, a lightweight Transformer modeling approach for reliable financial risk early warning. Using minute-level CSI 300 Index data as the research subject, we first fuse F-test and mutual information criteria to select 30 most discriminative risk-sensitive features, mitigating the distribution skew between historical training and live trading data. Second, we employ time-series causal partitioning (8:2) and rolling cross-validation to ensure evaluation robustness. Furthermore, model lightweighting and efficient edge deployment are achieved through Flatten-Patch embedding, single-head gated linear attention, and 8-bit dynamic quantization. Experiments demonstrate that Light-FinTS achieves R<sup>2</sup> = 0.970, MSE = 3232.5, and MAE = 43.7 on the test set, reducing MSE by approximately 90% compared to the standard Transformer. During periods of high market volatility, its direction prediction accuracy remains stable above 85%.The Wilcoxon signed-rank test (<i>p</i> &lt; 0.05) confirms the statistically significant performance improvement. Compared to recent lightweight models like DLinear and PatchTST, Light-FinTS better captures the local nonlinear dynamics of financial time series. After quantization, it achieves real-time inference within ≤ 50&#xa0;ms on a single-core CPU, significantly enhancing the responsiveness and deployment feasibility of early warning systems. This approach not only provides a practical technical pathway for high-frequency quantized strategies but also establishes a new paradigm for constructing trustworthy, efficient, and interpretable intelligent financial risk early warning systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A lightweight transformer modeling approach for trustworthy financial risk early warning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Mingxun Zhu,
  • Jiali Ma,
  • Yawen Tang,
  • Ying Wang

摘要

Early identification and reliable early warning of financial risks are critical challenges for maintaining market stability and serving financial security strategies. Addressing the limitations of existing models—including insufficient accuracy, weak interpretability, and high deployment costs under high-noise, strong nonlinearity, and low-latency constraints—this paper proposes Light-FinTS, a lightweight Transformer modeling approach for reliable financial risk early warning. Using minute-level CSI 300 Index data as the research subject, we first fuse F-test and mutual information criteria to select 30 most discriminative risk-sensitive features, mitigating the distribution skew between historical training and live trading data. Second, we employ time-series causal partitioning (8:2) and rolling cross-validation to ensure evaluation robustness. Furthermore, model lightweighting and efficient edge deployment are achieved through Flatten-Patch embedding, single-head gated linear attention, and 8-bit dynamic quantization. Experiments demonstrate that Light-FinTS achieves R2 = 0.970, MSE = 3232.5, and MAE = 43.7 on the test set, reducing MSE by approximately 90% compared to the standard Transformer. During periods of high market volatility, its direction prediction accuracy remains stable above 85%.The Wilcoxon signed-rank test (p < 0.05) confirms the statistically significant performance improvement. Compared to recent lightweight models like DLinear and PatchTST, Light-FinTS better captures the local nonlinear dynamics of financial time series. After quantization, it achieves real-time inference within ≤ 50 ms on a single-core CPU, significantly enhancing the responsiveness and deployment feasibility of early warning systems. This approach not only provides a practical technical pathway for high-frequency quantized strategies but also establishes a new paradigm for constructing trustworthy, efficient, and interpretable intelligent financial risk early warning systems.