FinSTaR: Towards Financial Reasoning with Time Series Reasoning Models

The article introduces FinSTaR (Financial Time Series Thinking and Reasoning), a model designed to enhance financial reasoning using a new benchmark called FinTSR-Bench, which consists of ten financial reasoning tasks based on S&P stocks. FinSTaR employs two distinct chain-of-thought strategies: Compute-in-CoT for deterministic assessments and Scenario-Aware CoT for stochastic predictions, achieving an average accuracy of 78.9%, significantly surpassing existing LLM and TSRM baselines. This development is crucial for practitioners as it provides a structured approach to financial time series analysis, addressing the unique challenges of the financial domain.