Research & Projects

Project 1

Construction Insurance & Industrial Risks Internship

I began this internship in April 2025 at BELOVETSKAYA AVOCATS, a Parisian law firm specialising in construction insurance, real estate law, and industrial risks. This was my first experience in a professional legal environment and was undertaken as part of the academic requirements for my Master 2 at Université Paris 1 Panthéon-Sorbonne. I worked under the supervision of Maître Daria Belovetskaya, with academic oversight from Professor Noufel Frikha. Credit goes to the full team at the firm for their guidance and mentorship throughout.

Objectives

The internship aimed to bridge my quantitative academic background with real-world legal and insurance practice. I sought to understand how liability is assessed in multiparty construction disputes, how insurers manage subrogation and claims recovery, and how financial risk is allocated through contract structures. An essential consideration throughout was learning to translate technical legal and financial documents into actionable insights for ongoing cases.

Results

Over the course of the internship I contributed to contractual analysis, drafting legal submissions, and claims management across active 20+ construction disputes. I gained hands-on experience with specialist legal tools including RPVA, ADAPPS, e-Barreau, and e-CARPA. A notable discovery was the dual role of the CCTP document, both as a pre-contractual specification and as retrospective evidence in litigation. The internship deepened my understanding of how quantitative risk assessment integrates with legal strategy, a combination I intend to carry forward into my career.

Project 2

Bayesian Approach to Excess of Loss Reinsurance Pricing

This research paper explores an improved pricing method for Excess of Loss (XOL) reinsurance, a critical tool insurers use to manage exposure to large claims. I pursued this project to address well-known limitations in traditional actuarial pricing methods, particularly their poor handling of incomplete data and long-tailed risk. The project was completed as part of the Master 2 in Risk Engineering: Finance and Insurance (IRFA) at Université Paris 1 Panthéon-Sorbonne.

Objectives

The core objective was to propose a Bayesian framework as a more reliable alternative to Maximum Likelihood Estimation (MLE) for XOL pricing. The initial challenge was that traditional methods often produce biased or unstable estimates when claims data is truncated or sparse. The project aimed to incorporate prior knowledge alongside observed data, account for emergence patterns in long-tailed lines of business, and produce more stable loss ratio estimates across reinsurance layers.

Results

Simulations showed that the Bayesian posterior loss ratio of 146.1% aligned more closely with the true simulated loss ratio of 133.4% than experience-based estimates, which overreacted to sparse data. The framework successfully reduced variance compared to MLE while maintaining acceptable bias. A key finding was that Bayesian methods are particularly valuable in high layers where reported claims are few. Future integration with Markov Chain Monte Carlo (MCMC) models could further enhance uncertainty quantification in real-world scenarios.

Project 3

Credit Risk Analysis : Loan Default Prediction

This data science project was completed as part of the Data Science Software course at Université Paris 1 Panthéon-Sorbonne, under the supervision of Bertrand Hassani. The project was a collaboration with Hoda Jaouhari (IRFA) and Iliass Maalim (MMMEF). We chose this topic because credit default prediction is a fundamental problem in financial risk management, and we wanted to apply machine learning techniques to a realistic, imbalanced financial dataset.

Objectives

The primary objective was to build a machine learning model capable of predicting whether a customer would default on a loan. The initial challenge was a significantly imbalanced dataset, where non-default cases far outnumbered defaults, a common real-world condition that can cause models to simply ignore the minority class. We sought to address this through oversampling, then evaluate multiple classification models to identify the most reliable approach.

Results

After applying random oversampling and standardising features, we trained and compared several models including Logistic Regression and Random Forest. The pipeline successfully handled the class imbalance and produced meaningful predictions on the test set. We found that tree-based models were better suited to capturing non-linear relationships in the data. The project reinforced the importance of data preprocessing and evaluation metrics beyond accuracy, such as precision and recall, when working with imbalanced financial data.

Page updated

Report abuse