Machine Learning and Econometrics for Risk Measurement in Finance – MLEforRisk

The MLEforRisk project is based on a methodological approach that tightly integrates financial econometrics and Machine Learning in order to renew the tools used to measure credit, market, and liquidity risk. The approach aims to move beyond the opposition between interpretable but restrictive econometric models and flexible but opaque ML algorithms by developing hybrid frameworks that combine predictive performance, interpretability, and statistical rigor, while remaining compatible with the operational and regulatory constraints of the financial sector.

In the area of credit risk, the methods rely on enriching standard econometric models with ML techniques designed to capture non linearities and complex interactions. This approach is implemented in particular through the integration of rules extracted from decision tree type models into penalized econometric frameworks, making it possible to improve predictive performance while preserving interpretability and statistical inference properties. The work also includes a systematic comparison of the predictive and economic performance of econometric and ML models in realistic credit scoring settings.

The methods then evolve toward an explicit consideration of algorithmic fairness issues. The project develops formal statistical tests to assess different definitions of fairness, as well as interpretability tools aimed at identifying the variables responsible for discriminatory biases. Post processing based correction procedures are also proposed to improve model fairness without significantly degrading predictive performance.

For market risk, the project relies on large scale multivariate models based on dynamic systems and conditional dependence structures. The approaches are built on parsimonious econometric frameworks and regularization techniques designed to make estimation feasible when the number of assets is large. These methods further evolve toward theoretical developments showing how low order multivariate systems can generate long memory behavior at the individual level, while improving risk forecasting.

In parallel, the project develops statistical validation procedures for systemic risk measures. In particular, it proposes backtesting tests for systemic risk measures that explicitly account for estimation risk.

Across all components, the project favors estimation strategies based on likelihood methods, penalized approaches, and Bayesian techniques, with particular attention paid to out of sample robustness, economic interpretability, and the statistical validity of the models developed.

The MLEforRisk project has helped to better understand how Machine Learning methods can be used in a reliable and responsible way to measure financial risk. A first key result is the demonstration that ML algorithms, when used on their own, can improve predictive accuracy but also exhibit important limitations in terms of interpretability, stability, and fairness. The project shows that these limitations can be overcome by combining ML with econometric methods, in order to build models that are both performant and interpretable.

The project’s work has led to new tools that make it possible to better explain why a risk model produces certain decisions, for example in the context of credit allocation. These approaches make models more transparent for users, financial institutions, and oversight bodies. The project has also developed methods to detect and correct potential biases in models, thereby helping to limit the risks of algorithmic discrimination and to strengthen the social acceptability of these tools.

Another important result concerns the improvement of statistical methods used to analyze the evolution of risk in financial markets. The project has made it possible to better account for market instability, complex dependencies between assets, and the occurrence of extreme events. These advances provide a more refined understanding of risk dynamics and improve the reliability of tools used to monitor financial tensions and anticipate phases of fragility in the financial system.

Finally, the project has had a notable impact on the dissemination of knowledge to both the public and decision makers. The results have been presented at international conferences, during workshops dedicated to early career researchers, and within institutional bodies involved in financial oversight. Through these activities, MLEforRisk has contributed to the public debate on the use of ML in finance and on the conditions under which these technologies can support a more reliable, transparent, and responsible financial system.

The results of the MLEforRisk project open up several scientific and applied perspectives in the medium and long term. From a methodological standpoint, the advances achieved in inference for complex dynamic models and in the controlled integration of ML within econometric frameworks provide a foundation for further developments. A first perspective concerns the extension of these approaches to higher dimensional environments, simultaneously incorporating temporal dependence, cross sectional dependence between assets, and unobserved heterogeneity, while preserving robustness and interpretability.

The work on model explainability and the decomposition of predictive performance also opens important perspectives for a better understanding of the mechanisms underlying algorithmic decisions. It calls for further research on model stability over time, the analysis of regime changes, and the early identification of weak signals announcing deteriorations in risk. These developments are particularly relevant in settings where models must be monitored, audited, and regularly recalibrated.

In the area of credit risk and algorithmic fairness, the methods developed can be extended to dynamic frameworks, multi source data, and other forms of automated decision making in finance. They pave the way for a joint analysis of economic performance, risk, and the social acceptability of models, in a changing regulatory environment.

The project’s contributions to the modeling of market risk and complex dependence structures also offer perspectives for improving tools used to monitor financial tensions and to enrich macro and micro prudential stress testing frameworks. Further developments are envisaged toward multi country and multi sector settings, as well as toward scenario analysis explicitly incorporating macroeconomic or financial shocks.

Finally, the project has structured a sustainable scientific dynamic at the interface between econometrics and ML, notably through the organization of the QFFE conference and workshops dedicated to early career researchers. This dynamic opens perspectives for new international collaborations, interdisciplinary projects, and advanced training initiatives, contributing to the long term development of these research themes in financial econometrics.

The growing use of artificial intelligence and Machine Learning (ML) by banks and Fintech companies is one of the most significant technological changes in the financial industry over past decades. These new technologies hold great promise for the future of financial services, but also raise new challenges. In this context, the MLEforRisk project aims to provide better understanding of the usefulness of combining econometrics and ML for financial risk measurement. This project aims to provide a rigorous study of the benefits and limitations of these two approaches in the field of risk management, which is the core business of the financial industry. MLEforRisk is a multidisciplinary project in the fields of finance and financial econometrics which brings together junior and senior researchers in management, economics, applied mathematics, and data science.
The project has five methodological objectives related to credit, market, and liquidity risks. In the context of credit risk, ML methods are known to provide good classification performances. However, these methods often black boxes, which is particularly problematic for both clients and regulators. Thus, our objective is to develop hybrid approaches to credit risk modeling by combining econometrics and ML to overcome the trade-off between interpretability and predictive performance. At the same time, the use of ML in the field of credit risk has led to a debate on the potential discrimination biases which could be generated by these algorithms. Here, our objective is to develop statistical methods to test the algorithmic fairness of credit risk models and to mitigate these biases.
In the area of market risk, the project aims to combine ML techniques and advanced econometric modeling to improve the estimation of conditional risk measures associated to portfolio returns. Our objective is to propose new hybrid approaches for modeling the conditional variance matrix of returns or its inverse, called the precision matrix. Since these risk measures are the key input of trading strategies, the accuracy of their estimation is essential for the asset management industry. These estimation methods will be designed in the perspective of large portfolios for which the number of assets can exceed by far the number of time observations available to estimate the moments. A second objective is to take into account the asymmetry of the conditional distribution of returns when modeling the conditional risk by using ML methods.
Concerning liquidity risk, we observe that the development of alternative market indices and factorial investment significantly modify the dynamics of traded volumes on the markets by increasing dependencies and network effects. Our objective is to take these effects into account when measuring liquidity risk, while reducing the dimension of the parameter set used in the network with ML methods.
The MLEforRisk project aims at creating a doctoral training network for young researchers specialized in financial econometrics. It also aims to promote a reproducible research. All codes and data produced within the project will be archived on RunMyCode and the reproducibility of the numerical results will be certified by cascad, the first certification agency for scientific code and data.