ExplainabilityTransparencyTrustworthinessq-fin.RM
A Time Series Approach to Explainability for Neural Nets with Applications to Risk-Management and Fraud Detection
Marc Wildi, Branka Hadji Misheva
December 6, 2022
arXiv.org
1 citations
Key Takeaways
-
Explainability is critical for financial AI adoption: The paper emphasizes that trust in AI systems for finance depends on understanding why models make specific predictions, not just accepting their outputs—a foundational principle for responsible deployment in high-stakes sectors.
-
Existing explainability methods have limitations for time series data: Classical XAI techniques work well for static datasets but struggle with time-ordered financial data that contains dependencies and non-stationary patterns, meaning current tools may provide misleading explanations for fraud detection and risk management systems.
-
Time ordering matters for financial decisions: The proposed approach preserves the natural sequence of events in financial data, which is essential because the timing and order of transactions fundamentally shape fraud patterns and risk—treating financial data as unordered could obscure the actual reasoning behind model decisions.
-
Specialized XAI techniques are needed for different data types: The paper demonstrates that one-size-fits-all explainability methods are insufficient; financial institutions need tailored approaches that account for the unique characteristics of their data to genuinely understand and trust their AI systems.
Abstract
Artificial intelligence is creating one of the biggest revolution across technology driven application fields. For the finance sector, it offers many opportunities for significant market innovation and yet broad adoption of AI systems heavily relies on our trust in their outputs. Trust in technology is enabled by understanding the rationale behind the predictions made. To this end, the concept of eXplainable AI emerged introducing a suite of techniques attempting to explain to users how complex models arrived at a certain decision. For cross-sectional data classical XAI approaches can lead to valuable insights about the models' inner workings, but these techniques generally cannot cope well with longitudinal data (time series) in the presence of dependence structure and non-stationarity. We here propose a novel XAI technique for deep learning methods which preserves and exploits the natural time ordering of the data.