When Reasoning Fails: Evaluating 'Thinking' LLMs for Stock Prediction
A 2025 paper by Sodha on machine learning return prediction in equities, using a large language model: the abstract mentions an out-of-sample test and transaction costs and leaves out a benchmark, multiple testing and the size of the universe.
“We compare a direct LLM (gpt-4o-mini), a TLLM (gpt-5), and classical learners (ridge, random forest) on cross-sectional ranking loss 1 - IC, MSE, and long/short backtests with realistic costs.”
AssetEquitiesindices
StrategyMachine learningmacro signals
TechniqueLarge language modeltree ensemble
GeographyIndia
Also stated: Long-short, out-of-sample test, costs
Missing from abstract: benchmark, multiple testing, universe
Institutional-grade audit of this strategy.
£5 — Request the audit- Audit of the paper claims accounting for costs, luck, crises, originality, geography, etc.
- Reproduction code integratable into Bloomberg, IBKR, etc.
- Bias-free datasets for backtesting included
- Delivered in 2 working days
The paper's neighbourhoodMachine learning return prediction by year
The family is counted across the whole universe. Asset, method and authors are counted inside the family.