Virtual Attendance – Bloomberg Quant (BBQ) September

In this seminar chaired by Bruno Dupire, Agostino Capponi will present the keynote, followed by "lightning talks" of 5 minutes each in quick succession. 5:30 PM Keynote: Agostino Capponi | Columbia University When More Data Hurts: Navigating the Limits of AI and Market Nonstationarity Does more training data always improve return forecasts? In non-stationary markets, longer windows expose models to structural change, while complex models that mitigate misspecification demand precisely such windows. We formalize this nonstationarity-complexity tradeoff: the excess risk of a model class decomposes into a misspecification term, an estimation-variance term governed by the class’s Rademacher complexity, and a drift term that grows with the training window under non-stationarity. We resolve it with a tournament-based selection procedure that adaptively chooses the optimal model class and training window using non-stationary validation data. We develop an oracle-type inequality guaranteeing performance close to the best model class-window pair in hindsight, uniformly over the unknown drift path. Applied to 17 industry portfolios, the method improves out-of-sample R^2 by 14–23% over rolling-window and regime-switching benchmarks, with gains concentrated in recessions. 6:30 PM Lightning Talks • Jessica Mitra | EDHEC Business School Does Accounting for Higher Moments Help Characteristic-Based Factors to Span the Stochastic Discount Factor? • Milind Sharma | QuantZ Capital / QMIT The Quantamental Revolution • Xiaotong Hu | John Hopkins University Synthetic Data for Time Series Foundation Models • Sasha Stoikov | Cornell University Finding Good Wines on the Cheap