Reading News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns

01 ABSTRACT

The study uses NoLBERT, an encoder pretrained only on 1976-1995 text and contrastively fine-tuned into a sentence encoder, to generate news topics and build attention shocks and stock exposures for forecasting returns of the 1,000 largest U.S. stocks. Models are fitted on 2013-2019, validated on 2020-2021, and tested on 2022-2024; the best validation model loses money at every rebalancing frequency, and extending the attention-shock reference window from 5 to 63 trading days raises test rank IC for all readers. NoLBERT produces the most stable topics and forecasts about as well as LDA; an equal-weight combination of all readers and models earns a monthly net Sharpe ratio of 1.12, significant against zero before search adjustment but not against a price-only benchmark.

02 KEY FINDINGS

1. The model with the best validation score loses money out-of-sample (2022-2024) at every rebalancing frequency, suggesting validation-based selection may not generalize.
2. Changing the attention-shock reference window from 5 trading days to the past 63 trading days raises test rank IC for all topic extractors (LDA, MiniLM, FinBERT, NoLBERT).
3. NoLBERT yields the most stable topics and forecasts about as well as LDA; the equal-weight combination of all models and readers has a monthly net Sharpe of 1.12 but is not significant relative to a price-only benchmark.