FinTradeBench: A Financial Reasoning Benchmark for LLMs

FinTradeBench is a new benchmark for testing whether large language models can make sound financial decisions by reasoning about both company fundamentals (from financial statements) and trading signals (from stock price patterns). The benchmark includes 1,400 questions about real NASDAQ companies and tests three types of reasoning: those focused only on fundamentals, only on trading signals, and hybrid questions requiring both. Testing 14 different language models reveals that while retrieving relevant documents helps with understanding company data, current AI models still struggle significantly with numerical reasoning and understanding time-series patterns in stock prices.