Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models

In one sentence

Summary

This paper addresses the challenge of training autonomous financial agents that generalize beyond specific market patterns, aiming to replace complex multi-agent systems with a single, locally deployable model capable of risk-aware reasoning. The authors argue that current open-source models lack the structured decision logic required for stable trading in noisy environments, motivating a focus on cognitive fine-tuning rather than simple prediction or sentiment analysis. The core contribution is a framework that combines high-quality data curation with a novel reasoning template to teach models how to form decisions under uncertainty, rather than just memorizing answers. The methodology centers on the Cognitive Financial Reasoning Dataset, constructed from classic textbooks and 15 years of S&P 500 historical data, verified by an AI committee of three frontier models to ensure label quality. The training pipeline employs CORA, a four-stage chain-of-thought template (Contextualize, Organize, Reason, Act), and DARA, a dual-axis augmentation strategy that shuffles answer positions and varies technical indicator parameters to prevent shortcut learning. The model, Llama-3.1-8B, is fine-tuned using Q-LoRA on this enriched data. Evaluation uses a two-stage protocol: Stage I measures static MCQ accuracy on a held-out test set, while Stage II assesses sequential trading performance through a chronological simulation of 150 episodes across bullish, bearish, and mixed regimes, tracking returns, exposure, and downside risk. Results show that the full model achieves 82.38% accuracy in Stage I, outperforming open-source baselines and approaching frontier models, while ablation studies confirm that CORA and DARA are critical for preventing degenerate behavior. In Stage II, the model delivers a 7.64% average return, significantly outperforming baselines and maintaining controlled downside risk in bullish and mixed regimes. However, the model exhibits a long-biased policy that underperforms in bearish markets, highlighting a limitation in regime adaptability. The study concludes that structured cognitive training enables smaller models to exhibit competitive, risk-aware trading behavior, though further work is needed to improve robustness across all market conditions.

Research problem

The paper investigates whether large language models can acquire generalizable financial decision-making competence that transcends specific market patterns, addressing the lack of structured training methods for autonomous trading agents. It specifically tackles the challenge of training models in noisy markets where ground truth is ambiguous, aiming to move beyond simple prediction or sentiment analysis toward robust, risk-aware reasoning. The authors also address the deployment risks of closed-source models by proposing a framework for locally deployable, open-weight models.

Core contributions

• Introduction of the Cognitive Financial Reasoning Dataset, a curated collection of multiple-choice questions derived from textbooks and historical market data, verified by an AI committee.
• Development of CORA (Contextualize–Organize–Reason–Act), a four-stage chain-of-thought template that structures financial reasoning into explicit, expert-like decision pathways.
• Proposal of DARA (Dual-Axis Robustness Augmentation), a data augmentation technique that mitigates shortcut learning by shuffling answer positions and varying technical indicator parameters.
• Design of a two-stage evaluation framework that combines static MCQ accuracy testing with a chronological, portfolio-constrained trading simulation to assess generalization to sequential decision-making.
• Demonstration that fine-tuning an open-source 8B model with this framework allows it to outperform open-source baselines and approach frontier-model performance in specific market regimes.

Method

• Data Construction: Extracted MCQs from 10 textbooks and 15 years of S&P 500 data, using GPT-4o for summarization and question generation.
• Verification: Used an AI committee of three models to verify labels, with human review for cases with no consensus.
• CORA Template: Expanded each MCQ into four reasoning steps: Contextualize, Organize, Reason, and Act.
• DARA Augmentation: Applied 24 permutations of answer choices and varied technical indicator parameters to reduce shortcut learning.
• Fine-Tuning: Used Q-LoRA to fine-tune Llama-3.1-8B-Instruct on the augmented dataset with cross-entropy loss on assistant tokens.
• Stage I Evaluation: Measured accuracy on a held-out test set of 210 MCQs with deterministic decoding.
• Stage II Evaluation: Simulated trading over 150 episodes, tracking portfolio value, exposure, and turnover under realistic constraints.

Results & findings

• The full model achieves 82.38% accuracy in Stage I, outperforming all open-source baselines (e.g., Qwen3-8B at 71.90%) and approaching frontier models.
• Ablation studies show that removing CORA causes accuracy to collapse to 46.19%, indicating the critical role of structured reasoning traces.
• Removing DARA reduces accuracy to 76.67% and degrades trading performance, highlighting the importance of robustness augmentation.
• In Stage II, the full model achieves a 7.64% average return, significantly outperforming open-source baselines like Llama-3.1-8B (6.23%).
• The model exhibits strong performance in bullish regimes (43.66% return) and mixed regimes, with controlled downside risk (CVaR-5% of -2.64% in mixed).
• The model underperforms in bearish regimes, suggesting a long-biased policy that is not well-adapted to sustained downturns.
• High static accuracy is necessary but not sufficient for trading success; the CORA/DARA pipeline helps translate reasoning into consistent portfolio decisions.

Data & experiments

• Cognitive Financial Reasoning Dataset: 1,400 initial MCQs from 10 textbooks and 15 years of S&P 500 data, expanded to ~27,800 training examples via DARA.
• Verification: An AI committee of three frontier models (Gemini-2.5-Flash, Claude-3.7-Sonnet, Llama-4-Scout) verified labels, with human review for disagreements.
• Stage I Experiment: Static evaluation on a held-out test set of 210 MCQs to measure isolated reasoning accuracy.
• Stage II Experiment: Chronological trading simulation with 150 non-overlapping episodes (50 per regime: bullish, bearish, mixed) using S&P 500 data.
• Metrics: Stage I uses accuracy; Stage II uses average return, mean net exposure, turnover, and CVaR-5% (downside risk).
• Baselines: Frontier models (GPT-5.1, Claude-4.5-Sonnet, DeepSeek-v3.1) and open-source models (Llama-3.1-8B, Qwen3-8B, Mistral-7B).

Limitations & caveats

• The model exhibits a long-biased policy that underperforms in bearish regimes, limiting its generalizability across all market conditions.
• The evaluation uses frictionless execution, ignoring real-world trading frictions like bid-ask spreads and slippage.
• The dataset is limited to S&P 500 data and technical indicators, potentially lacking diversity in asset classes and fundamental signals.
• The AI verification committee may still contain biases or errors, although human review was used for disagreements.
• The study focuses on a single 8B model, and the generalizability to other model sizes or architectures is not fully explored.

Curator notes

Discovery status: round1_promoted_seed_for_round2. Source links: AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework; BizBench: A Quantitative Reasoning Benchmark for Business and Finance; BloombergGPT: A Large Language Model for Finance; ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering; DISC-FinLLM: A Chinese Financial Large Language Model based on Multiple Experts Fine-tuning; FinBERT : A Large Language Model for Extracting Information from Financial Text†; FinBERT: Finan