Revisiting the Risk Elicitation Puzzle: A Machine Learning Approach
Project Overview
This project tests whether risk preferences transfer across elicitation methods. Using Zhou & Hey's four-instrument panel (Holt–Laury price lists, pairwise lotteries, certainty equivalents, and convex budget allocations, 96 subjects), I train encoder–decoder models to predict a person's choices on one method given their behaviour on the others, scored against a per-problem crowd floor with person-level cross-validation. The original question was dimensional: if transfer exists, how many latent dimensions carry it? To make a null interpretable, I built a simulator on the real problem designs with planted cross-method-correlated preferences, so the same pipeline can be shown to detect signal when signal exists.
What were the key results of your research project?
A precisely measured null. On real data, cross-method transfer is indistinguishable from zero after crediting per-problem crowd rates. Crucially it's a measurement, not a failure. The same pipeline, folds and floors detect planted signal strongly on simulated data, so "no signal" and "broken model" are separable.
Pooling architectures fail, retrieval ones don't. Every model that compresses a person into a fixed-width vector lands below the crowd floor; the cross-attention model that keeps per-problem structure reaches positive skill on sim. Diagnosis conditioning collapse, the decoder learns base rates from the query task alone and has little gradient pressure to use the person representation.
A presentation of this research will be shared here and on our YouTube site when available.
How do you feel you have benefitted from completing this internship and has it made you consider future career paths?
I learnt how to build models that cater to not just this specific problem but any problem where the output is linked to many features. Training a model on a dataset was a huge problem that I had to tackle early on. Spotting problems like conditioning collapse, as well as coming up with new ideas and architectures in a field where I had no prior experience, was also very rewarding and gave me confidence in my foundations.
I want to take this project further to write a paper with my supervisor since the findings are valuable and we have more to contribute.
Download presentation slides