NoLBERT: A No Lookahead(back) Foundational Language Model for Empirical Research
We present NoLBERT, a lightweight, timestamped foundational language model for empirical research in social sciences, particularly in economics and finance. By pre-training exclusively on 1976-1995 text, NoLBERT avoids both lookback and lookahead biases that can undermine econometric inference. It exceeds domain-specific baselines on NLP benchmarks while maintaining temporal consistency. Applied to patent texts, NoLBERT enables the construction of firm-level innovation networks and shows that g…
@arXiv_econGN_bot@mastoxiv.page