Language model topics from past news predict stock returns without hindsight
Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns
Computation and LanguageMachine Learning
Summary
Financial language models often get an unfair advantage by training on the same years they are tested on, giving away future information. The authors trained a model called NoLBERT only on news texts from 1976 to 1995 and tested it on news from 2012 to 2024 to avoid this. They compared it to other models in predicting stock returns and found that NoLBERT created the most stable topics but performed similarly to an older method called LDA. Combining all models gave modest predictive results that were not better than just using past prices.
What this means in practice
- •For quantitative investment teams: Use stable topics from historically trained language models as an additional signal when forecasting stock returns without future data leakage.
- •For financial data engineers: Implement contrastive fine-tuning on historical financial news text to generate interpretable topic features that avoid hindsight bias in return prediction models.
Authors
Kevin Foley, Jonathan Hartadi, Shivesh Prakash, Charishma Takkallapalli, Swapnil Vatsal
Abstract
Language models that read financial news are usually trained on text from the same years used to test them, so a backtest can reward hindsight. We build news topics from NoLBERT, an encoder pretrained only on text published between 1976 and 1995, adapted into a sentence encoder by contrastive fine-tuning on passages of the same article, with no labels or returns. We compare it with LDA, MiniLM and FinBERT readers on 1.26 million articles from 2012 to 2024, convert each reader's topics into attention shocks and stock exposures, and forecast the returns of the 1,000 largest US stocks with eight models at three rebalancing frequencies. Models are fitted on 2013-2019, choices are made on 2020-2021, and 2022-2024 is held out. The experiment with the best validation score loses money at every frequency. Measuring attention shocks against the past 63 trading days instead of 5 raises the test rank IC of every reader. NoLBERT gives 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 adjusting for the search but not against a price-only benchmark.