In-Context Learning of Price Impact: Toward Optimal Execution from a Handful of Trades
Sound trading decisions require a model of price impact: how much a trade moves the price against the trader and how long that effect persists. Yet price impact is hard to measure and even harder to predict. It varies across stocks and market regimes, and each measurement requires a costly trade that itself moves the market.
In-Context Operator Networks (ICON) offer an approach to this challenge. ICON is a transformer pre-trained on simulated trade and price trajectories drawn from a prior over price impact models. Given a handful of trades and their associated price responses as context, ICON infers the prevailing price impact operator without retraining. When used as a surrogate impact operator in an optimal execution problem, it yields near-optimal trading strategies for the regime inferred from the in-context examples.
We demonstrate this approach using linear and nonlinear propagator models with path-dependent transient price impact, in which the effect of a trade persists and decays over time according to a propagator kernel. These models admit computable optimal execution strategies, providing a benchmark for evaluating the learned operator. We also investigate fine-tuning the pre-trained network using real order flow from TAQ market data to explain out-of-sample intraday price paths.
The talk is based on ongoing joint work with Georg Menz, Nils Detering, Tingwei Meng, Giulio Farolfi, and Stanley Osher.
Bio: Moritz Voss is a faculty member in the Department of Mathematics at UCLA, where he serves as Associate Director of Financial Mathematics and is affiliated with the Mathematical and Statistical Finance (MSF) Research Group. Before joining UCLA, he was a Visiting Assistant Professor in the Department of Statistics and Applied Probability at UC Santa Barbara. He earned his Ph.D. in Mathematics from TU Berlin in 2017 under the supervision of Peter Bank and holds a master’s degree in Probability and Finance (DEA El Karoui) from Université Pierre et Marie Curie (Paris 6). His research focuses on optimal stochastic control problems in algorithmic trading, including order execution, alpha trading, optimal tracking, and market making, with an emphasis on price impact modeling and market microstructure. His work combines stochastic analysis, machine learning, and the analysis of high-frequency limit order book data.


