In briefPoint-in-Time Encoders for Earnings Calls: Text Vectors Without Look-Ahead Bias
SummaryThis paper trains text encoders on earnings call transcripts in yearly vintages from 2010 to 2024, where each vintage reads only text dated on or before its cutoff to avoid look-ahead bias in backtests. The models are trained on 412,248 calls and fine-tuned on call question-answer pairs. Vintages trained only on calls retrieve answers better than the general retrievers tested, and their vectors raise the rank correlation between next-day returns and forecasts. Point-in-time vintages matter most for retrieval and offer little gain for return forecasts.
In briefDeep learning approximation of operator maps in optimal trade liquidation
SummaryThis paper studies an optimal execution problem where prices follow dynamics affected by a general transient impact kernel. It shows that the optimal trading strategy depends locally Lipschitz continuously on the model parameters governing temporary and transient price impact and the underlying price process, which justifies neural network approximations using these parameters as inputs. The authors propose methods for generating kernel samples, projecting the kernel to a finite-dimensional representation, and reducing dimension based on the relation between inventory and trade signal. They use the trained network to quantify potential financial costs of misspecifying the transient impact kernel.
Ryan Donnelly, Luca Galimberti, Brian Healy et al.
