Founded in 1991, we are a global quantitative and systematic asset management firm applying a scientific approach to finance to develop alternative investment strategies that create value for our clients.
We value innovation, dedication, collaboration, and the ability to make an impact. Together, we create a stimulating environment for talented and passionate experts in research, technology, and business to explore new ideas and challenge existing assumptions.

Summary

Build and stress-test generative models based on flow matching/stochastic interpolants for synthetic time-series. You will create controllable synthetic datasets (with tunable noise, nonlinearity, regime changes, and exogenous controls) and/or adapt benchmarks from the literature, compare neural architectures (MLPs, Transformers, etc.) and training strategies, and quantify:

1. Sample complexity and data efficiency across architectures
2. Whether learning a full conditional distribution improves accuracy when we only care about conditional means
3. Out-of-support generalization with control variables (how far can we extrapolate?)

Motivation

Diffusion-style generative models (flow matching / stochastic interpolants) work remarkably well for images and text, but their behavior on structured time-series—especially with controls and mixed scalar/sequence inputs—is less well understood. Also, in realistic settings we are in a limited data regime, which is not encountered in vision or NLP. This project asks: which architectures and training choices are most data-efficient, when is generative modeling worth it versus simple regressions, and how robust are these models to shifts in inputs statistics?

These are the main questions of interest; the internship will prioritize a subset based on progress and the intern’s interests.

Prerequisites: comfort with Python/PyTorch and basic probability/ML. Curiosity about generative modeling and careful experimental design.

We are continuously striving to be an equal opportunity employer and we prohibit any discrimination based on sex, disability, origin, sexual orientation, gender identity, age, race, or religion. We believe that our diversity, breadth of experience, and multiple points of view are among the leading factors in our success.
CFM is a signatory of the Women Empowerment Principles.
 

Follow us on Twitter or LinkedIn or visit our website to find out more about CFM.

Profile description:

Prerequisites: comfort with Python/PyTorch and basic probability/ML. Curiosity about generative modeling and careful experimental design.