This event is free and open to all (RSVP). You will receive the webinar link from no-reply@zoom.us upon registration.
AI With Error Bars
We love GenAI, LLMs, and the like. But there currently are no estimates of sampling error in its output, no error bars. The result harkens back to the fictional computer Deep Thought*, which stated that the answer to life, the universe and everything was 42. In this talk, we seek a world where 42 has a confidence interval that can yield to useful quant work.
We focus on portfolio optimization of a group of 15 stocks to explore the effect of finite sample sizes, and the sample variance it produces in weights and performance measures. It covers both classical and machine learning approaches to construction. That is, the mean-variance (MV) method pioneered by Markowitz, and the heirarchical risk parity (HRP) method of Lopez.
The known strong sensitivity in MV to portfolio weights underlines the importance of qualifying their sampling distribution, and enables us to compute error bars and confidence intervals on performance and other measures. In the case of HRP, we show that the world of AI can actually be as precise as the freshman physics experiment of a swinging pendulum with known period 2 Pi Sqrt(L/g).
*Douglas Adams in”The Hitchhiker’s Guide to the Galaxy”
Bio: Jonathan Schachter is CEO of Delta Vega, Inc., a financial engineering consultancy in Brooklyn. He is lead author of “A First Textbook in Model Validation and Model Risk Management” (Academic Press, January 2026). A 25-year veteran Wall Street quant, Jonathan hosts regular Special Topics in Risk Management (STORM) talks. He also trains financial engineering students at Columbia and Stony Brook Universities. He is a market expert at P.R.I.M.E. Finance, which assists judicial systems in the resolution of disputes concerning complex financial transactions.