Scaling Point-in-Time Language Models

Bryan Kelly, Semyon Malamud, Johannes Schwab +1

Large language models trained on unrestricted internet corpora inevitably embed information from the future, introducing lookahead bias that compromises the validity of backtests a…