Reproduce “DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yutong Yan, Raphael Tang, Zhenyu Gao, Wenxi Jiang, Yao Lu
DatedGPT trains twelve 1.3B-parameter vintages from 2013–2024, filtering pretraining data by Common Crawl crawl date and applying year-specific instruction tuning. On 61,290 firm-day headlines from 2014–2024, prior-year vintages produce a lookahead-free signal worth 28.3 basis points per standard deviation and a 3.20 Sharpe ratio. A later-vintage residual adds 26.4 basis points, but crawl dates imperfectly capture publication timing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.