FinSTaR: A Chain-of-Thought Strategy Tailored for Financial Temporal Reasoning

This paper proposes the FinSTaR financial temporal reasoning model. To address the characteristics of financial data—intertwined determinism and randomness, and the complexity of single-entity vs. multi-entity analysis—it constructs a 2×2 capability classification framework and adopts differentiated chain-of-thought strategies (Compute-in-CoT for deterministic tasks, Scenario-Aware CoT for random tasks). The model achieves an average accuracy of 78.9% on the FinTSR-Bench benchmark, significantly outperforming existing LLM and TSRM baselines.

Financial temporal reasoning faces twofold challenges: First, deterministic evaluation tasks (e.g., volatility calculation) and random prediction tasks (e.g., stock price trends) have large differences in nature, and existing models use the same method leading to suboptimal performance; Second, single-entity (single asset) and multi-entity (asset linkage) analysis have different complexities. To this end, the study proposes a 2×2 capability classification framework:

	Single-entity Analysis	Multi-entity Analysis
Deterministic Evaluation	Individual Indicator Calculation	Relative Performance Comparison
Random Prediction	Individual Trend Prediction	Portfolio/Relationship Prediction

The framework covers four capability quadrants, each requiring different reasoning strategies.

FinSTaR adopts differentiated chain-of-thought strategies for different tasks:

1. Compute-in-CoT (Deterministic Tasks)：Through programmatic reasoning of identifying required calculations → extracting data → step-by-step calculation → verifying results, ensuring accuracy and interpretability;
2. Scenario-Aware CoT (Random Tasks)：Generating multiple scenarios → evaluating scenario probabilities → reasoning within scenarios → comprehensive judgment, simulating analysts' scenario-based analysis to improve prediction rationality.

The study constructs the FinTSR-Bench benchmark, which includes 10 financial reasoning tasks (5 deterministic evaluations: technical indicator identification, etc.; 5 random predictions: direction prediction, etc.). Experiments show that FinSTaR achieves an average accuracy of 78.9%, significantly outperforming general-purpose LLMs (e.g., GPT-4) and general TSRMs. In addition, joint training of the four capabilities yields better results than separate training, reflecting complementarity; Scenario-Aware CoT outperforms standard CoT in all prediction tasks.

The implications of FinSTaR: Need to distinguish between types of financial tasks, and emphasize interpretability and scenario thinking. Application scenarios include: intelligent investment research assistants (assisting data analysis), risk management systems (scenario-based risk assessment), investment education tools (interpretable teaching), and regulatory technology (abnormal transaction identification).

Existing limitations: Data scope is limited to S&P stocks, time granularity is mainly daily, no integration of external information (e.g., news), and focus on correlation rather than causal reasoning. Future directions: Expand data scope and time granularity, integrate external information, and enhance causal reasoning capabilities.

FinSTaR provides a systematic framework and method for financial temporal reasoning, promoting the construction of AI systems that understand financial logic. The code has been open-sourced: https://github.com/seunghan96/FinSTaR.