Advanced Analytics and Artificial Intelligence, Intern

Context

Predictive modeling is central to an insurers mission: understanding assessing and predicting biometric Life & Health this often means working with survival analysis a domain that brings specific methodological challenges due to censored data non-proportional hazards complex multivariate relationships and the need for medically coherent outputs.

Within our team we have developed a robust Python-based survival modeling framework that adapts both traditional actuarial methods and modern machine-learning (ML) algorithms to censored data. This internal library already integrates a wide range of models from classical Cox variants to advanced ML approaches and continues to evolve to meet emerging needs such as handling richer datasets improving interpretability and aligning with new regulatory and actuarial standards.

To strengthen this foundation we have identified several R&D topics that will be the focus of the internship. Validated methods and results will be integrated into our existing survival modeling package.

This internship offers an opportunity to meet the requirements for an Actuarial Science degree while:

• Working in an international environment and collaborating with diverse teams
• Exploring cuttingedge topics at the intersection of AI and actuarial science
• Learning technical methodological and industry best practices

• Being mentored and supported by experts in the field

• Deploying solutions that deliver strong business value

Mission & R&D topics

Examples of topics of interest include:

• Integrating domain-driven constraints into our modeling framework
• Interpretability of survival models

In actuarial practice domain expertise often dictates how certain variables should influence risk. This creates the need to explore approaches that embed model constraintssuch as monotonicity convexity or U-shaped effectsdirectly into survival models. These constraints ensure that model outputs remain aligned with established medical and actuarial knowledge (for example enforcing that mortality risk follows a U-shaped relationship with BMI).

Explainable AI is a recurring priority in survival modeling.

Existing tools such as partial dependence plots (PDPs) accumulated local effects (ALE) and SHAP-like approaches for censored data offer valuable insights but also present key challenges. For example:

• PDPs may generate unrealistic feature combinations when inputs are correlated.

• Survival-specific SHAP variants remain computationally costly and sometimes unstable.

The goal is to investigate how these techniques can be improved or adapted to deliver more robust realistic and domain-consistent interpretations.

• Model Assessment & Diagnostics

• Model performance evaluation:

  A wide range of survival metrics exist each capturing different aspects of performance (calibration ranking consistency bias). This diversity can lead to:

  • Conflicting conclusions between metrics
  • Difficulty comparing models objectively
  • Challenges summarizing results into a single decision criterion.

  Research questions include:

  • Which metrics should be favored under which modeling context

  • How can we reconcile metrics

  • Is it possible to derive an aggregated performance score that synthesizes several evaluation angles

• Business alignment and domain consistency:

  Beyond statistical performance models must behave consistently with actuarial and medical expertise. A key area of R&D will be to develop a tool that:

  • Highlight model limitations especially where predictions contradict well-established risk patterns.
  • Provide diagnostics aligned with underwriting reasoning (e.g. assessing the isolated and combined effect of key drivers in a medically coherent way).
  • Identify profiles that deviate from expected behavior even in the presence of continuous risk factors where the space of possible profiles is theoretically infinite.

  This requires exploring systematic and exhaustive approaches to surface unexpected behaviors for example detecting monotonicity violations abrupt / discontinuous predictions or implausible interaction effects across the full covariate space.

• Testing new survival models
• Establishing best practices around various topics

The goal is to identify methods powerful enough to capture non-linearities and interactions yet less prone to overfitting and offering greater control than fully flexible ML models.

Potential avenues include Penalized and constrained Cox variants or interaction-augmented Cox models.

In addition testandlearn experimentation on less common and more specialized approaches such as causal survival forests could also be conducted

• Handling correlated input features
• Model transferability across portfolios or markets
• Turning model predictions into business outputs (e.g.- optimal risk grouping strategies)

Internship structure

• Phase 1

You will begin by reviewing:

• Our internal survival modeling library and its architecture

• Existing R&D work and technical documentation

• Relevant academic and actuarial literature.

This requires familiarity with Python object-oriented programming and you will learn or reinforce skills in unit testing documentation and development workflows.

• Phase 2

You will then take ownership of one or more R&D topics producing:

• New model components or methodological enhancements
• Implementation in Python following our coding standards
• Validation notebooks benchmarks and documentation
• A final research article and internship report suitable for an Actuarial Science thesis

Applicants Profile

• Ultimate or penultimate Master student in the following fields: Computer science Mathematics Biostatistics or Statistics.

• Strong interest in actuarial science and machine learning
• Experience with survival models is a major asset
• Solid Python skills (OOP scientific libraries) ;
• Curious rigorous and comfortable communicating insights to both

The internship may be completed over six months as a oneyear internship or during a gap year.