Explainable Artificial Intelligence (XAI) is increasingly required in computational economics, where machine-learning forecasters can outperform classical econometric models but remain difficult to audit and use for policy. This survey reviews and organizes the growing literature on XAI for economic time series, where autocorrelation, non-stationarity, seasonality, mixed frequencies, and regime shifts can make standard explanation techniques unreliable or economically implausible. We propose a taxonomy that classifies methods by (i) explanation mechanism: propagation-based approaches (e.g., Integrated Gradients, Layer-wise Relevance Propagation), perturbation and game-theoretic attribution (e.g., permutation importance, LIME, SHAP), and function-based global tools (e.g., Accumulated Local Effects); (ii) time-series compatibility, including preservation of temporal dependence, stability over time, and respect for data-generating constraints. We synthesize time-series-specific adaptations such as vector- and window-based formulations (e.g., Vector SHAP, WindowSHAP) that reduce lag fragmentation and computational cost while improving interpretability. We also connect explainability to causal inference and policy analysis through interventional attributions (Causal Shapley values) and constrained counterfactual reasoning. Finally, we discuss intrinsically interpretable architectures (notably attention-based transformers) and provide guidance for decision-grade applications such as nowcasting, stress testing, and regime monitoring, emphasizing attribution uncertainty and explanation dynamics as indicators of structural change.
The integration of Machine Learning (ML) and Deep Learning (DL) methods into economic and financial modeling has triggered a fundamental paradigm shift (Mullainathan & Spiess, 2017;Varian, 2014). Historically, econometrics has relied on linear models, such as multiple linear regression or Vector Autoregressions (VAR) (Lütkepohl, 2005), where transparency is inherent: a coefficient β provides a direct and global view of marginal effects (a unit increase in interest rates results in a decrease of β in investment, ceteris paribus) (Wooldridge, 2010). However, the rigidity of these models often fails to capture the complex nonlinear dynamics and higher-order interactions characteristic of modern financial markets and global supply chains (Koop et al., 1996).
In contrast, “black-box” models such as Recurrent Neural Networks (RNNs) or Transformers have demonstrated superior capacity to approximate these complex functions (Makridakis et al., 2018;Zerveas et al., 2021), but at the cost of obscuring the decision boundary (Rudin, 2019). This trade-off between accuracy and transparency has become unsustainable in the face of strict regulatory frameworks, such as the “Right to Explanation” under the General Data Protection Regulation (GDPR) in Europe (Goodman & Flaxman, 2017) or Basel III/IV risk management guidelines, which require financial institutions to justify capital requirements and credit decisions. Consequently, Explainable Artificial Intelligence (XAI) has emerged not as a luxury, but as a strategic and legal necessity (Barredo Arrieta et al., 2020), aiming to bridge the cognitive gap between the predictive power of algorithms and the human need for causal understanding (Doshi-Velez & Kim, 2017).
The challenge of applying XAI to economics lies in the idiosyncratic nature of economic time series. Unlike static image or tabular data, economic series exhibit autocorrelation, stochastic trends, seasonality, and structural regime shifts (Hamilton, 1994;Stock, 2020). Standard explainability methods that assume feature independence can disrupt the temporal structure of the data (Molnar, 2025), creating unrealistic “counterfactual” scenarios that violate the arrow of time or the coherence of the economic state (Tonekaboni et al., 2019). Recent literature has therefore begun to adapt general XAI techniques to the specific constraints of economic inference (Belle & Papantonis, 2021).
Within the broad spectrum of explainability techniques, propagation-based methods represent an “intrinsic” or white-box approach, leveraging the internal model architecture to trace the flow of information.
Propagation techniques, mainly used in Artificial Neural Networks (ANNs), exploit the differentiability and multiplicative composition of network layers to compute feature importance by tracking contributions from the model output back to its input. This approach is based on the theoretical premise that the final prediction can be decomposed into the sum of the relevance of the input neurons.
A foundational method in this category is Integrated Gradients (Sundararajan et al., 2017), which accumulates gradients along a linear path from a baseline to the current input. This method satisfies the completeness axiom, ensuring that the sum of attributions equals exactly the difference between the model prediction and the baseline prediction a crucial property for financial auditing, where each basis point change in a risk prediction must be justified.
Among the most representative and sophisticated methods is Layer-Wise Relevance Propagation (LRP) (Bach et al., 2015). Unlike raw gradients, which can be noisy or suffer from saturation (where changes in input no longer affect output due to the nature of activation functions), LRP propagates relevance scores layer by layer using conservation rules, redistributing evidence of the prediction backward.
In finance, applying LRP has enabled significant advances in understanding asset valuation models. Traditional factor models (such as the Fama-French three-factor model) assume static linear relationships between asset returns and risk factors like size, value, or momentum. However, Nakagawa et al. (2019) challenged this orthodoxy by applying LRP to a Deep Factor Network.
Their study used LRP to evaluate the dynamic contribution of multiple factors (risk, value, size, among others) in a neural network model trained to predict stock returns. Results revealed substantial divergences compared to traditional linear sensitivity measures, such as Kendall or Spearman correlations. While linear correlations suggested a constant exposure to certain factors, LRP analysis showed that the neural network captured nonlinear, regimedependent exposures. For instance, the “Value” factor could have a dominant positive influence during economic recovery periods but become irrelevant or negative during liquidity crises. This ability of LRP to visualize factor importance locally and temporally allowed the model
This content is AI-processed based on open access ArXiv data.
