Analogy-Driven Financial Chain-of-Thought

• The paper introduces AD-FCoT, a prompt-only framework that integrates analogical reasoning with step-by-step causal analysis for financial news sentiment prediction.
• AD-FCoT employs two analogical exemplars as historical causal templates, overcoming limitations of traditional lexicon-based and generic LLM prompts.
• Experimental results show AD-FCoT achieves superior precision (57.45%) and enhanced market correlation compared to zero-shot, few-shot, and other CoT methods.

Analogy-Driven Financial Chain-of-Thought (AD-FCoT) is a prompting framework for financial news sentiment analysis that integrates analogical reasoning with chain-of-thought prompting for sentiment prediction on historical financial news. It guides LLMs to draw parallels between new events and relevant historical scenarios with known outcomes, and to embed those analogies in a structured causal rationale before emitting a final Positive, Negative, or Neutral label. The method is explicitly prompt-only: it requires no additional training data or fine-tuning, and relies on analogical exemplars placed in context as causal templates for interpreting the target article (Singhal, 16 Sep 2025).

1. Intellectual setting and problem formulation

AD-FCoT is motivated by a specific diagnosis of financial news sentiment analysis. The relevant task is not generic polarity detection, but prediction of whether a news article’s effect on a company’s stock is Positive, Negative, or Neutral. The method’s starting point is that traditional lexicon-based and standard ML systems often fail to capture nuanced financial language and economic context, while naive prompting with general LLMs can produce surface-level sentiment judgments that do not model the underlying economic mechanism. The paper also argues that standard CoT can still miss the appropriate causal template when an event requires financial contextualization, and contrasts its approach with DK-CoT, which adds domain knowledge but does not explicitly exploit historical precedent via analogy (Singhal, 16 Sep 2025).

The central conceptual claim is that financial interpretation is often precedent-based. Product recalls, strong earnings reports, and regulatory changes matter not because of surface wording alone, but because they instantiate recurrent causal patterns with familiar market consequences. AD-FCoT therefore formalizes sentiment analysis as a case in which a model should ask what happened, why it matters economically, what historical pattern it resembles, and how that resemblance supports the final sentiment label. The paper presents this as, to the authors’ knowledge, among the first approaches in finance to explicitly combine analogical reasoning with CoT prompting (Singhal, 16 Sep 2025).

The paper also provides an information-theoretic motivation for why news-derived sentiment matters for market prediction. Let $Y$ denote short-term future stock price change or return, and $X$ denote information derived from news, such as a sentiment score. The uncertainty reduction supplied by news is written as

$I(X;Y) = H(Y) - H(Y \mid X).$

This formalism is motivational rather than algorithmic, but it clarifies the paper’s view that financial news sentiment should be evaluated in relation to market outcomes rather than only textual labels (Singhal, 16 Sep 2025).

2. Prompt architecture and reasoning process

AD-FCoT is defined as a prompting strategy with four essential elements: a financial-analyst role instruction, a small number of analogical few-shot exemplars, an explicit chain-of-thought cue, and the target article. The paper states the prompt structure as instruction + 2 analogical examples + causal reasoning cue + target article. It also gives an example task instruction: “You are a financial analyst. Read the news and reason step-by-step about its impact on the company’s stock, then output Positive/Negative/Neutral.” The analogical examples are selected from a historical events pool and each contains a news snippet, a causal reasoning chain, and a sentiment label (Singhal, 16 Sep 2025).

Prompt component	Function
Instruction / Task Description	Set analyst role and output space
Two analogical exemplars	Supply historical-like causal templates
Causal reasoning cue	Encourage step-by-step grounded analysis
Target news query	Present the article to classify

The analogies are not represented as lexical matches but as historically familiar financial scenarios whose causal structure resembles the target event. The paper’s examples include a product recall as a historically bad scenario and a strong earnings report as a historically good scenario. In this design, the exemplars do not merely provide answers; they show the reasoning pattern the model should imitate. The intended chain includes identification of the event, analysis of its economic impact, relation to analogous historical scenarios, inference of stock sentiment, and final label output (Singhal, 16 Sep 2025).

A crucial architectural limit is that AD-FCoT does not include automated analogy retrieval. The paper does not specify embedding retrieval, nearest-neighbor search, external historical database lookup, or similarity scoring. It states only that the prompt contains two analogical exemplars selected from a historical pool, and that prompt engineering and exemplar selection use pre-2023 data. This means the method’s analogy mechanism is in-context and static rather than retrieval-augmented or dynamically case-based (Singhal, 16 Sep 2025).

3. Experimental protocol and reported performance

The evaluation uses the FNSPID dataset, described as containing thousands of labeled financial news articles spanning 2012–2023 and covering S&P 500 companies. The split is explicitly temporal: data through 2022 are used for prompt engineering and exemplar selection, while January–June 2023 are reserved as the test set. The paper presents this split as a protection against look-ahead bias, especially because the LLaMA pretraining cutoff is stated as the end of 2021 (Singhal, 16 Sep 2025).

All experiments use Meta LLaMA 3, 8B, untuned. The system runs on the Groq AI accelerator stack with the HuggingFace Transformers API in BF16 precision on a private server with four GroqCards, and reported latency is under 100 ms/token. Prompt length is kept under 1024 tokens, output generation is capped at 256 tokens, and decoding uses $T=0$. The compared prompting conditions are Zero-Shot, Few-Shot, CoT, DK-CoT, and AD-FCoT (Singhal, 16 Sep 2025).

The main reported metrics are Accuracy, Precision, and Recall, with the usual definitions:

$$\text{Accuracy} = \frac{\text{TP} + \text{TN}}{\text{TP} + \text{FP} + \text{TN} + \text{FN}},$$

$$\text{Precision} = \frac{\text{TP}}{\text{TP} + \text{FP}},$$

$$\text{Recall} = \frac{\text{TP}}{\text{TP} + \text{FN}}.$$

The paper interprets these in trading terms as overall directional correctness, correctness of positive signals, and capture rate of actual up-moves, respectively (Singhal, 16 Sep 2025).

Method	Accuracy	Precision	Recall
Zero-Shot	53.92	44.95	48.80
Few-Shot	54.70	54.11	51.42
CoT	51.81	54.27	50.20
DK-CoT	52.09	55.62	53.45
AD-FCoT	54.92	57.45	53.62

AD-FCoT therefore reports the best values on all three metrics, though the margin over Few-Shot in Accuracy is only 0.22 percentage points. The paper emphasizes the larger gain in Precision, where AD-FCoT reaches 57.45 versus 55.62 for DK-CoT. It also claims substantially higher correlation with market returns, but the provided text does not include exact correlation coefficients or a correlation table. A qualitative example states that, on a product recall article, AD-FCoT referenced a similar past recall and correctly inferred negative impact, whereas baseline CoT missed the appropriate causal template (Singhal, 16 Sep 2025).

The paper also notes an evaluation ambiguity. The nominal task is ternary sentiment classification, but the evaluation discussion says neutral price changes below tick-size threshold are discarded because they do not trigger a trading decision. The paper does not fully reconcile this tension between ternary labeling and effectively more directional evaluation (Singhal, 16 Sep 2025).

4. Position within financial reasoning research

AD-FCoT belongs to a broader line of work on structured financial reasoning, but it occupies a distinct niche: analogical prompting for news sentiment analysis. Related systems emphasize different control primitives. FinCoT grounds reasoning in expert financial workflows encoded as Mermaid blueprints rather than analogical exemplars, and reports improvements such as 63.2% to 80.5% on Qwen3-8B-Base while also reducing generated tokens relative to structured CoT prompting (Nitarach et al., 19 Jun 2025). FINDER combines generative retrieval with dynamic in-context example selection and Program-of-Thought execution for numerical financial QA, and its dynamic exemplar selection behaves as an instance-conditioned analogy mechanism even though the paper does not call it that (Khatuya et al., 15 Oct 2025).

FinSTaR addresses a different axis of financial reasoning by separating deterministic assessment from stochastic prediction, using Compute-in-CoT for the former and Scenario-Aware CoT for the latter. That distinction is highly relevant to AD-FCoT because analogical reasoning appears naturally suited to uncertain forecasting tasks rather than to tasks that are exactly computable from observed prices (Lee et al., 5 May 2026). Financial-PoT reaches a similar conclusion from the arithmetic side: it argues that semantic understanding and deterministic computation should be architecturally decoupled, showing that standard CoT suffers from “Arithmetic Hallucinations” and “Cognitive Collapse” on complex quantitative tasks (Zhao et al., 29 Jan 2026).

System	Primary setting	Main control mechanism
AD-FCoT	Financial news sentiment	Analogical exemplars + CoT
FinCoT	CFA-style financial QA	Expert workflow blueprints
FINDER	Financial numerical reasoning	Dynamic exemplars + PoT
FinSTaR	Financial time-series reasoning	Compute-in-CoT / Scenario-Aware CoT

Other adjacent works illuminate parts of the AD-FCoT design space without implementing analogy directly. FinChain supplies symbolic, executable financial traces and the ChainEval metric for step-level evaluation, which suggests a route for evaluating whether an analogy-derived chain is structurally correct rather than merely plausible (Xie et al., 3 Jun 2025). AlphaFin and Stock-Chain show that financial CoT combined with retrieval over reports, news, and market data can improve usefulness and backtested return, but their retrieval module is evidence-oriented rather than precedent-oriented (Li et al., 2024). LongFinanceQA and PAI decompose long-context reasoning into property extraction, property-conditioned retrieval, and summarization, which suggests a way to align financial entities and periods before any analogical comparison is attempted (Lin et al., 18 Feb 2025).

5. Limitations, unresolved issues, and common misconceptions

A common misconception is that AD-FCoT performs historical-case retrieval. It does not. The analogies are prompt-level exemplars chosen from a historical pool, not dynamically retrieved at inference time, and the paper explicitly leaves automated analogy retrieval to future work (Singhal, 16 Sep 2025). A second misconception is that AD-FCoT is a trainable architecture; it is not. It is a prompt-only method with no additional training data or fine-tuning (Singhal, 16 Sep 2025).

The paper also acknowledges or implies several methodological limitations. There is no formal ablation isolating which AD-FCoT component drives the gains, and the authors explicitly state that such analysis is future work. The method likely depends on exemplar selection, but the stability of that choice is not systematically studied. The approach also remains dependent on the underlying model’s internal financial knowledge and reasoning ability, because there is no external historical memory, no retrieval module, and no structured mechanism for testing whether an analogy is genuinely applicable to the target event (Singhal, 16 Sep 2025).

Further limitations concern evaluation and interpretability. The paper states that AD-FCoT explanations align with domain expertise and mirror human analytical reasoning, but it does not report a separate human-scored rationale benchmark. It also claims stronger correlation with market returns without providing exact correlation values in the supplied text. The benchmark itself uses same-day price direction as a proxy for sentiment, and the paper notes that market reaction can diverge from textual tone; negative news can be received positively if outcomes are better than feared. That caveat constrains what “sentiment improvement” means economically (Singhal, 16 Sep 2025).

A broader controversy, suggested by adjacent work, is whether analogy alone is sufficient for robust financial reasoning. Results from Financial-PoT indicate that even when semantic understanding is good, arithmetic-heavy tasks still require deterministic execution (Zhao et al., 29 Jan 2026). Results from FinSTaR indicate that deterministic assessment and stochastic prediction should not be treated with a single reasoning style (Lee et al., 5 May 2026). These findings do not invalidate AD-FCoT, but they suggest that analogy is most naturally a mechanism for schema selection and uncertain-event interpretation, not a universal replacement for expert workflow design or programmatic computation.

6. Research trajectory and likely extensions

The most plausible extensions of AD-FCoT are visible in the surrounding literature. Personalized Chain-of-Thought Summarization of Financial News for Investor Decision Support introduces a four-stage pipeline from raw articles to keyword-conditioned investor actions, with intermediate event-focused summaries and metadata extraction. This suggests that AD-FCoT could benefit from an explicit intermediate representation layer rather than reasoning directly over raw article text (Zhang et al., 24 Oct 2025). Agentar-DeepFinance-300K argues that financial reasoning quality depends heavily on the quality, depth, and correction of synthesized CoTs, which suggests that analogy-driven reasoning may require curated, high-quality analogical traces rather than only better prompts (Zhao et al., 17 Jul 2025).

A plausible implication is that a more mature AD-FCoT system would combine several components that are currently distributed across different papers. From FinCoT, it could inherit expert workflow structure; from PAI, property extraction and decomposed retrieval over long financial documents; from FINDER, dynamic in-context example selection and executable reasoning; from FinChain, step-level verifiability; and from Financial-PoT, strict decoupling between semantic planning and numerical execution (Nitarach et al., 19 Jun 2025, Lin et al., 18 Feb 2025, Khatuya et al., 15 Oct 2025, Xie et al., 3 Jun 2025, Zhao et al., 29 Jan 2026).

This suggests an emerging interpretation of AD-FCoT not as a finished method class but as a research agenda. In that agenda, analogies are most useful for selecting a causal template, aligning current events with prior financial scenarios, and explaining why a precedent transfers or fails to transfer. Deterministic subproblems would still be delegated to programmatic or verifiable reasoning, while prediction problems under uncertainty would use analogy as a structured form of historical precedent. On that reading, the original AD-FCoT paper establishes the prompt-level prototype—historical exemplars plus causal CoT—and the adjacent literature indicates how it could evolve into a retrieval-grounded, schema-aware, and execution-backed financial reasoning framework (Singhal, 16 Sep 2025).