Seven Sins of Investing with GenAI

Research evidence excerpt

Seven Sins of Investing with GenAI

not be unit-tested at generation time. The harness must therefore substitute point-in-time

evidence construction, source hierarchy, citation gates, evaluator gates, and release controls for immediate

pass/fail feedback.

The bias literature is especially important because GenAI investing expands the classic quantitative research

failure modes. Kong, et al [2026] argue that LLM evaluation in finance requires explicit bias treatment, including

temporal leakage, survivorship, objective design, narrative bias, and cost bias. Model leakage and look-ahead bias

are the most severe. In traditional empirical research, look-ahead bias usually enters through restated data, revised

series, or future universe membership (i.e., survivorship bias). In GenAI workflows, future information may be

embedded in pretrained parameters, retrieval corpora, post-event narratives, source metadata, or the model's

ability to infer company identity from context. Lopez-Lira, Tang, and Zhu [2025] show that LLMs can memorize

realized economic outcomes, while Gao, Jiang, and Yan [2026] provide direct tests of look-ahead bias in LLM

forecasts. The practical implication is that point-in-time discipline must cover the model checkpoint, prompt,

retrieval layer, document corpus, and output release process, not only the input document date.

Hallucination and storytelling bias form another channel of contamination. General NLP work has developed

several useful defenses: retrieval-augmented generation (RAG) grounds answers in external documents, retrieval-

augmented refinement (RARR) revises generated claims against retrieved evidence, Chain-of-Verification

separates answer generation from independent checking, and SelfCheckGPT uses response inconsistency to flag