When Reasoning Fails: Evaluating 'Thinking' LLMs for Stock Prediction

Abstract

Problem. "Thinking" LLMs (TLLMs) expose explicit or hidden reasoning traces and are widely believed to generalize better on complex tasks than direct LLMs. Whether this promise carries to noisy, heavy-tailed and regime-switching financial data remains unclear. Approach. Using Indian equities (NIFTY constituents), we run a rolling 48m/1m walk-forward evaluation at horizon k = 1 day and dial cross-sectional complexity via the universe size U in {5, 11, 21, 36} while keeping the reasoning budget fixed (B = 512 tokens) for the TLLM. 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. Statistical confidence is measured with Diebold-Mariano, Pesaran-Timmermann, and SPA tests. Main findings. (i) As U grows under a fixed budget B, the TLLM's ranking quality deteriorates, whereas the direct LLM remains flat and classical baselines are stable. (ii) TLLM variance is higher, requiring ex-post calibration (winsorization and blending) for stability. (iii) Portfolio results under transaction costs do not support a net advantage for the TLLM. Hypotheses. Our results are consistent with the following testable hypotheses: H1 (Capacity-Complexity Mismatch): for fixed B, TLLM accuracy degrades superlinearly in cross-sectional complexity. H2 (Reasoning Variance): TLLM outputs exhibit higher dispersion date-by-date than direct LLMs, increasing error bars and turnover. H3 (Domain Misfit): next-token prediction objectives and token-budgeted inference are poorly aligned with heavy-tailed, weakly predictable stock returns. Implication. In our setting, "thinking" LLMs are not yet ready to replace classical or direct methods for short-horizon stock ranking; scaling the reasoning budget and/or re-aligning objectives appears necessary.