The Marginal Yield of LLM-Guided Alpha Search: Search Effort, Information and Investment Value

Language models can propose stock-return factors without limit, so the useful question is what an additional proposal is worth to an investor. We follow the gain from continued search through three stages: the training score the search optimizes, the accuracy of a joint return forecast, and the net return of a long-only portfolio. Paired experiments vary the search policy, the information available to the search and the downstream learner. The gain is stage-dependent. Training scores improve quickly and then saturate; price-volume search raises its best score by 17% within twelve further proposals. Little of this carries through. Repairing diagnosed formula defects raises factor scores while lowering forecast accuracy, and no policy changes net return by an amount distinguishable from zero. Tracing the libraries through the pipeline shows where the gains thin out: new factors overlap more with features the model already uses than the factors they replace, re-estimation offsets part of their contribution, and the stocks they add earn about what the displaced stocks earned, so transaction costs are not the cause. Research effort on factor search is best judged by the return it delivers, not by the score it optimizes.