InvestorBench: A Benchmark for LLM Financial Decision-Making Tasks

InvestorBench: A Benchmark for LLM Financial Decision-Making Tasks Figure 1: Schematic diagram of the overall architecture of the InvestorBench framework. In addition, as application scenarios expand from a single target to multiple financial tasks, the influencing factors in financial decision-making will also become complex and diverse depending on the type of tasks. For example, individual stock trading needs to focus on company- and industry-level fundamentals and financial report data (Yi et al. (2022)); while cryptocurrency trading is more likely to be driven by real-time news and market sentiment (Bhatnagar et al. (2023)); as for ETFs, they often focus on long-term asset allocation and passive investment strategies (Madhavan (2016)). Although a series of LLM agent frameworks targeting a single financial scenario have recently appeared, such as FinMem (Yu et al. (2024a)), FinAgent (Zhang et al. (2024a)), CryptoTrade (Li et al. (2024)), FinRobot (Yang et al. (2024)) and FinCon (Yu et al. (2024b)), etc., but these solutions often only focus on a single or a few financial tasks, and most of them rely on private data, making it difficult to conduct horizontal comparative evaluations. Therefore, there is an urgent need for an open benchmark for multi-task scenarios that can not only uniformly…