Replication Code and Data for 'Chronologically Consistent Large Language Models'

Published: 12 August 2026| Version 1 | DOI: 10.17632/s6yt64mrwb.1
Contributors:
Songrun He, Linying Lv, Asaf Manela, Jimmy Wu

Description

This dataset contains the data, code, and documentation for replicating the results in “Chronologically Consistent Large Language Models,” forthcoming in the Journal of Financial Economics. The package reproduces all tables and figures from the included derived series and provides the code for the underlying asset-pricing and language-model pipelines. Dow Jones Newswires and CRSP are proprietary and therefore are not redistributed; synthetic stand-ins with identical schemas are provided so that the data-processing and asset-pricing pipeline can be run end-to-end without access to the proprietary data. The package also documents the public pretraining corpora and released ChronoBERT and ChronoGPT model checkpoints. See README.md for data provenance, computational requirements, and detailed reproduction instructions.

Files

Steps to reproduce

Detailed reproduction instructions are provided in README.md. To reproduce all tables and figures from the shipped derived data: pip install -r requirements/requirements_analysis.txt ./run_all.sh exhibits To verify that the complete data-processing and asset-pricing pipeline runs end to end using the included synthetic substitutes for Dow Jones Newswires and CRSP: ./run_all.sh smoke To reproduce the analysis from licensed Dow Jones Newswires and CRSP data, follow Section 3 of the README, set the required environment variables, and run: ./run_all.sh full Model-pretraining and chronological-consistency reproduction procedures, software environments, hardware requirements, and expected outputs are documented in README.md and the README files within the relevant code directories.

Categories

Finance, Artificial Intelligence

Licence