A flagship fund with a return curve that points almost only upwards. The task for everyone else in the seminar was simple: find out why you shouldn’t believe this curve.
The seminar was Biases and Mishaps in Statistics and Machine Learning, held in English, with 15 people. Everyone had to hand in two things, both graded: a talk and a case study the others work through themselves.
I called my case study Project Midas. Behind it is a synthetic dataset I generated with a Python script: 60 stocks over 120 months and 1,000 backtests of trading strategies. That’s where the fund’s nice curve comes from. And it contains three mistakes that happen in real financial data all the time.
First, survivorship bias. If you only look at the stocks that survived to the end, you automatically see better returns. Second, look-ahead leakage. At one point a strategy uses next month’s data to decide today. An off-by-one error, one line, and suddenly the strategy can see the future. Third, multiple testing. Out of 1,000 tested strategies, a few look great by pure chance. Show only those, stay quiet about the rest, and you have a fund that looks too good. On top of that come a few decoys, oddities that are harmless, so people can’t just flag everything that looks strange.
The group got a handout, slides, an answer key and a walkthrough of the solution. Every number in there can be regenerated from the script. That mattered to me. If I want to teach people not to trust numbers blindly, mine should be checkable. For the multiple-testing part I drew on a 2014 paper by Bailey, Borwein, López de Prado and Zhu.
The talk on May 20 was the second piece: 30 minutes on hallucination in language models, then 15 minutes of discussion. I opened with Mata v. Avianca, the case where a lawyer cited court decisions in a filing that a chatbot had made up. Then the theory: how hallucination is defined, measured, detected and reduced, for example with semantic entropy, Self-RAG or RLHF. And a result by Kalai and Vempala saying that calibrated language models must hallucinate. At the end, the bridge to statistics: the prosecutor’s fallacy and base-rate neglect, meaning why even a good detector can raise mostly false alarms when the thing it looks for is rare. I read about 30 papers for it.
Both pieces came with an AI disclosure log, a record of where I used AI and for what. That was required, and it fits the topic.
What I’m taking away: at work I build products on language models myself. If you don’t plan for hallucination there, you find out from the customer. And hiding a mistake on purpose is harder than finding one. It has to be subtle, but findable.
Tools and skills: Python with pandas, NumPy and matplotlib, Slidev for the slides, literature research, statistics around backtesting, giving a scientific talk in English.
Note added later: the module grade is a 1.3.