DC16: Investigating the utility of classical XAI methods in financial time series

Objectives

The introduction of complex ML and DL methods for financial time series forecasts potentially enables higher predictive accuracy but this comes at the cost of higher complexity and thus lower interpretability. 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. The literature currently does not offer any XAI approach that is specifically developed for financial time series. Further research is needed on developing explainability methods that can be applied to complex models like deep learning methods (DL) which preserve and exploit the natural time ordering of the data.

Expected Results

Within this IRP, we will propose a set of novel explainability functions that are specifically tailored for financial time series. We envision a framework for XAI in finance that addresses the shortcomings of existing methods. Namely, under existing, perturbation-based XAI methods, if features are correlated, the artificial coalitions created will lie outside of the multivariate joint distribution of the data. Furthermore, generating artificial data points through random replacement disregards the time sequence hence producing unrealistic values for the feature of interest. In addition to the novel, finance-tailored methodology for obtaining explanations, the project will also aim to produce industry-ready deployments of the novel XAI techniques developed.

Secondments (3)

Recruitment & Hosting Details

DC 16 BFH53, 54 University of Twente M9 36

17

DC 16 BFH UTW Month 9 36 months D 3.1 - 3.3

Investigating the utility of classical XAI methods in financial time series (WP 3)

Objectives: The introduction of complex ML and DL methods for financial time series forecasts potentially enables higher predictive

accuracy but this comes at the cost of higher complexity and thus lower interpretability. 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. The literature currently does not offer any XAI approach that is

specifically developed for financial time series. Further research is needed on developing explainability methods that can be applied to

complex models like deep learning methods (DL) which preserve and exploit the natural time ordering of the data.

Expected Results: Within this IRP, we will propose a set of novel explainability functions that are specifically tailored for financial time

series. We envision a framework for XAI in finance that addresses the shortcomings of existing methods. Namely, under existing,

perturbation-based XAI methods, if features are correlated, the artificial coalitions created will lie outside of the multivariate joint

distribution of the data. Furthermore, generating artificial data points through random replacement disregards the time sequence hence

producing unrealistic values for the feature of interest. In addition to the novel, finance-tailored methodology for obtaining explanations,

the project will also aim to produce industry-ready deployments of the novel XAI techniques developed.

Planned secondments: ECB, Dr. Lukaz Kubicki, M21, 12 months, exposure to globally leading central bank research, training on

EU principles.

Fraunhofer, Prof. Dr. Ralf Korn, M33, 6 months. Research needs to be validated with industry to achieve the envisioned impact,

BIS, Rafael Schmidt, M39, 6 months, contribute macro-economic datasets, ongoing projects as well as overall expertise in banking

supervision

Fellow Host institution PhD enrolment Start date Duration Deliverables

17

Deliverables