Machine Learning Researcher

The problem.

A learned model on financial data is mostly an efficient way to overfit. The question is not whether a model can be trained. It is whether the full training procedure — every choice made while building it — would still produce something on untouched data. You own that question and the machinery that answers it.

What you would do.

• Build training and evaluation pipelines where split, embargo and feature construction cannot leak, and where leakage is caught by a test rather than a reviewer.
• Compare against a simple baseline honestly. Be willing to report that the baseline wins.
• Take a model from a notebook to something that loads, runs inside a strict latency budget, and fails safe on malformed input.
• Decide what the system does when a model cannot be fitted at all. That is a real state, not an edge case.
• Keep the count of everything tried, including architectures and hyperparameters. Those are trials like any other.

What we would want evidence of.

• You have caught leakage in your own pipeline and can describe how it got there.
• You can explain the calibration of a model's output, not only its accuracy.
• You have deployed a model into a system where being wrong costs money, and know what you changed afterwards.
• You are sceptical of your own architecture choices in proportion to how much you like them.
• You can work in a strictly typed codebase and a compiled inference path without asking for a private research environment.

What would make this a bad fit.

This role is a poor fit if your interest is model architecture for its own sake. Most of the work is evaluation, data and deployment.

This role is a poor fit if you need a large cluster and a large dataset before you can start. The data is small enough to think about and large enough to fool you.