PALM: Point-in-Time Adaptation for Financial Language Models

PALM replaces annual pretraining for financial language models with low-rank adaptation (LoRA) to prevent look-ahead bias in backtests. The method avoids full retraining cycles by fitting adapters to chronologically filtered corpora, matching the performance of yearly checkpoints without the high computational cost of complete pretraining runs.