Date: September 15, 2025
Time: 3:45-5 p.m.
Location: Gates 310 or via Zoom
Speaker: Maxwell (Max) Fishelson, Ph.D. candidate in the Theory of Computation group at MIT
Abstract: A set of probabilistic forecasts for a binary outcome is calibrated if each forecast closely approximates the empirical distribution. For example, looking back at all days a weatherperson forecasts 20% rain, we should see rain on about 20% of those days. The classic work of Foster and Vohra (1998) demonstrates that it's actually possible to produce calibrated forecasts on average over time, even if the outcomes are selected by an adversary. In this talk, we present an improved algorithm that guarantees calibrated forecasts with even less time, discuss lower bounds that rule out substantially faster calibration, and connect the whole thing to a fun combinatorial game that illuminates the structure of the problem.
Joint work with: Yuval Dagan, Constantinos Daskalakis, Noah Golowich, Bobby Kleinberg, Princewill Okoroafor
https://arxiv.org/abs/2406.13668
Bio: My name is Maxwell (Max) Fishelson. I'm a final year Ph.D. candidate in the Theory of Computation group at MIT. I’m on the faculty and postdoc job market for the 2025-2026 cycle. My research spans learning theory and algorithmic game theory. I focus on online learning in games and high-dimensional settings, with emphasis on calibration and swap-regret benchmarks.