FinCoT: Grounding Chain-of-Thought in Expert Financial Reasoning

This paper presents FinCoT, a structured chain-of- thought (CoT) prompting approach that incorporates insights from domain-specific expert financial reasoning to guide the reasoning traces of large language models. We investigate that there are three main prompting styles in FinNLP: (1) standard prompting—zero shot prompting; (2) unstructured CoT—CoT prompting without an explicit reasoning structure, such as the use of tags; and (3) structured CoT prompting—CoT prompting with explicit instructions or examples that define structured reasoning steps. Previously, FinNLP has primarily focused on prompt engineering with either standard or unstructured CoT prompting. However, structured CoT prompting has received limited attention in prior work. Furthermore, the design of reasoning structures in structured CoT prompting is often based on heuristics from non-domain experts. In this study, we investigate each prompting approach in FinNLP. We evaluate the three main prompting styles and FinCoT on CFA-style questions spanning ten financial domains. We observe that FinCoT improves performance from 63.2% to 80.5% and Qwen-2.5-7B-Instruct from 69.7% to 74.2%, while reducing generated tokens eight-fold compared to structured CoT prompting. Our findings show that domain-aligned structured prompts not only improve performance and reduce inference costs but also yield more interpretable and expert-aligned reasoning traces.

In finance this can lead to omissions in valuation checks or confusion between basis points and percentages. Yet real-world analysis follows well-defined workflows-valuation, discounting, portfolio attribution-that depend on explicit intermediate structure. Embedding such structure in the prompt helps the model verify units, formulas, and boundary conditions, improving interpretability and alignment with expert practice.

taxonomy of financial prompting– covering standard prompting, unstructured CoT, and structured CoT/FinCoT

Building upon ST-CoT approaches, FinCoT explicitly embeds expert-derived problemsolving methodologies directly into prompts, guiding LLMs to follow domain-specific reasoning pathways without requiring model fine-tuning.

3. Expert Reasoning Blueprint (via Mermaid Diagram (Sveidqvist and contributors, 2025):

A domain-specific, embedded expert blueprint with Mermaid diagram concerning the generation of diagrams (see §4), serve as a "Hint"within the context of the prompt. This blueprint explicitly outlines the recommended step-bystep problem-solving strategy for the given financial domain

creation of effective expert reasoning blueprints involves a systematic multistage process designed to transform a wide range of financial knowledge into structured and actionable LLM diagrams.

Curation Pipeline for Expert Reasoning: To construct expert blueprints for each financial domain, we implement a hybrid pipeline combining machine assistance and human judgment, as illustrated in Figure 3. The process includes the following stages:

1. Scope Definition and Knowledge Aggregation: The target CFA domain is scoped (e.g., Economics focusing on supply and demand), and relevant expert strategies are aggregated, using Deep Research2, from diverse authoritative sources. Outputs are validated by humanin- the-loop reviewers with financial knowledge (e.g., finance graduates or CFA charterholders) to ensure conceptual accuracy and domain alignment.

2. Validation and Synthesis: We cross-reference and synthesize the aggregated knowledge to ensure accuracy, identify core principles and filter redundancies.

3. Iterative Refinement into StructuredWorkflows: The synthesized expert knowledge is iteratively transformed into logical step-by-step reasoning workflows. This refinement process focuses on ensuring the coherence, correctness, and clarity of the resulting problem-solving strategies for each financial domain.

4. Mermaid Diagram Generation: This refined workflow is translated into a Mermaid diagram (Bari et al., 2024) using its text-based syntax. We selected Mermaid due to its LLM prompt compatibility and clear visual guidance aligning with the FinCoT prompt. The diagrams are constructed based on the source content validated and synthesized first in 2, and then applied to each financial domain, with the entire collection organized and described in Appendix A as reference blueprints.

5. Prompt Integration: The text-based Mermaid blueprint is embedded as “Hint" within the Fin- CoT prompt template (Appendix B.2), directly guiding the LLM’s reasoning.