The paper introduces Look-Ahead-Bench, a standardized benchmark designed to measure look-ahead bias in Large Language Models (LLMs) used for financial trading. Unlike existing benchmarks that rely on question-answering tasks, Look-Ahead-Bench evaluates models within a realistic agentic trading workflow based on the AI Hedge Fund framework, which makes actual portfolio decisions. The benchmark employs a dual-period design: an in-sample period (April–September 2021) within the training window of standard LLMs, and an out-of-sample period (July–December 2024) beyond their knowledge cutoff. The key metric is alpha decay, defined as the change in excess return over a buy-and-hold benchmark between the two periods. Negative alpha decay indicates look-ahead bias, as the model's relative performance deteriorates when tested on genuinely new data.

The authors evaluate several open-source standard LLMs (Llama 3.1 8B and 70B, DeepSeek 3.2) and a family of Point-in-Time (PiT) models (Pitinf-Small, Pitinf-Medium, Pitinf-Large) designed to eliminate look-ahead bias by having a strict knowledge cutoff of January 2020. Results show that all standard LLMs exhibit significant negative alpha decay (ranging from -15.25 to -21.77 percentage points), indicating strong look-ahead bias. In contrast, Pitinf models show positive alpha decay, with performance improving from in-sample to out-of-sample. Notably, the paper identifies a 'Scaling Paradox': larger standard models suffer more severe decay due to stronger memorization of training data, while larger Pitinf models benefit from improved reasoning capabilities. The work establishes a foundation for standardized evaluation of temporal bias in financial LLMs and provides a practical tool for identifying models suitable for real-world deployment.