Backtesting a trading strategy that reads financial news usually involves a hidden flaw: the language model has already seen the news from the test period during training. That overlap means the model's apparent success may come from hindsight rather than genuine prediction. The authors of a new arXiv paper argue this problem is widespread and propose a remedy.

Their solution is NoLBERT, an encoder pretrained only on text published before the evaluation window. By building news topics from this strictly historical corpus, the researchers aim to measure how news topics relate to the cross-section of stock returns in a way that a trader could have actually used in real time.

The paper's contribution is methodological as much as empirical. It offers a way to evaluate news-based signals without the contamination that plagues many existing backtests, and it demonstrates how such topics can be constructed and applied. The abstract does not report specific return results, so the main takeaway is the design principle: train only on the past, then test on the future.