Scholarlink

<p>The proliferation of Artificial Intelligence (AI) in financial forecasting necessitates rigorous scrutiny beyond predictive accuracy, particularly concerning model behavior during market turmoil. While Explainable AI (XAI) aims to demystify “black box” models, the consistency of a model’s reasoning process under stress—defined here as <b>explainability stability</b>—remains largely unexplored. This paper proposes and implements a novel framework combining systematic, data-driven anomaly detection (Isolation Forest filtered by VIX criteria) with comparative SHAP analysis to evaluate both the performance robustness and explainability stability of distinct machine learning archetypes (Ridge, RandomForest, XGBoost) forecasting Nasdaq Composite returns (2015–2025). We introduce quantitative stability metrics, validated by rigorous <b>placebo testing</b>, to distinguish genuine regime-dependent instability from statistical artifacts of data scarcity. Our findings reveal a critical decoupling between predictive resilience and interpretive consistency. While Ridge regression demonstrates the highest stability, RandomForest exhibits severe interpretive volatility (Mean Absolute Relative Change &gt; 670%) despite robust predictive performance—a phenomenon confirmed by placebo tests to be driven by market stress rather than sample size. XGBoost displays hypersensitivity to volatility inputs. This study underscores that performance robustness does not guarantee stable explanations, advocating for the integration of dynamic stability metrics into the standard validation toolkit for trustworthy financial AI.</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

Robustness Vs. Explainability Stability: A Comparative XAI Analysis of Machine Learning Models in Financial Market Anomalies

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Mohammadreza Ayatollahi,
  • Seyed Mohammadbagher Jafari

摘要

The proliferation of Artificial Intelligence (AI) in financial forecasting necessitates rigorous scrutiny beyond predictive accuracy, particularly concerning model behavior during market turmoil. While Explainable AI (XAI) aims to demystify “black box” models, the consistency of a model’s reasoning process under stress—defined here as explainability stability—remains largely unexplored. This paper proposes and implements a novel framework combining systematic, data-driven anomaly detection (Isolation Forest filtered by VIX criteria) with comparative SHAP analysis to evaluate both the performance robustness and explainability stability of distinct machine learning archetypes (Ridge, RandomForest, XGBoost) forecasting Nasdaq Composite returns (2015–2025). We introduce quantitative stability metrics, validated by rigorous placebo testing, to distinguish genuine regime-dependent instability from statistical artifacts of data scarcity. Our findings reveal a critical decoupling between predictive resilience and interpretive consistency. While Ridge regression demonstrates the highest stability, RandomForest exhibits severe interpretive volatility (Mean Absolute Relative Change > 670%) despite robust predictive performance—a phenomenon confirmed by placebo tests to be driven by market stress rather than sample size. XGBoost displays hypersensitivity to volatility inputs. This study underscores that performance robustness does not guarantee stable explanations, advocating for the integration of dynamic stability metrics into the standard validation toolkit for trustworthy financial AI.