Artificial intelligence in risk management: explainability and causality as key pillars

What it is about

How artificial intelligence is transforming risk management, the importance of explainable AI, and the implementation of causal models for decision-making in complex and regulated environments.

The webinar reviews the limitations and inconsistencies of widely used explanation methods, such as SHAP and LIME values, with examples of common applications, and applies causal models and explanation methods that incorporate the explicit causality of the variables.

It includes a case study in which explainability and causality were key to solving challenges in the financial sector, and shows how these tools can be used to anticipate and mitigate risks.

Session summary

The speaker explains why explainability and causality are pillars for using artificial intelligence in risk management. He reviews seven types of use cases (pattern recognition, event detection, forecasting, personalization, interaction support, optimization and reasoning with knowledge structures), the levels of interaction between people and systems, from augmented intelligence to autonomous systems, and the need to always have a responsible person. He argues that an artificial intelligence governance scheme must include ethics, transparency, privacy and non-discrimination.

He illustrates with examples that correlation is not causation: the correlation between stocks and bonds changes over time, Simpson’s paradox shows how omitting a variable flips the sign of a relationship, and a model that classified dogs and wolves was guided by the snow in the background. He presents explainable artificial intelligence techniques, inherently interpretable models versus post-hoc tools such as LIME and SHAP, local or global, and a LIME variant with causality. He shows a probability-of-default case in which SHAP reveals that the postal code carries enough weight to create bias, and a study of financial crises and growth. In the questions he cautioned that there is no specific regulation yet and that these methods are not infallible.

Topics covered

• Seven types of artificial intelligence use cases in finance
• Levels of interaction between people and systems: from augmented to autonomous intelligence
• Artificial intelligence governance: ethics, transparency, privacy and accountability
• Correlation and causation: examples and confounding variables
• Explainable artificial intelligence: LIME, SHAP and local and global explanations
• Cases: probability of default, postal-code bias and financial crises
• Questions: regulation and limits of explanation methods