FinCARE: Financial Causal Analysis with Reasoning and Evidence

In one sentence

Summary

Portfolio managers often rely on correlation-based heuristics that fail to capture true causal drivers of performance. FinCARE addresses this by proposing a hybrid framework that augments statistical causal discovery with domain knowledge from SEC 10-K filings and LLM reasoning. The system extracts causal triplets from filings to build a knowledge graph, which provides algorithmic constraints for causal discovery algorithms. Simultaneously, an LLM module generates hypotheses for missing edges based on financial theory. This approach aims to ground statistical discoveries in financial expertise while maintaining empirical validation, enabling proactive risk management and strategic decision-making in dynamic markets. The methodology enhances three causal discovery paradigms: constraint-based (PC), score-based (GES), and continuous optimization (NOTEARS). Knowledge graph constraints are encoded as edge weights, with required edges bypassing independence tests in PC and adding bonuses to the BIC score in GES. In NOTEARS, constraints are applied via regularization and post-processing. The LLM component uses a 'MissingEdgeDiscoverer' to propose edges, which are integrated similarly to KG constraints. Experiments utilize a synthetic dataset of 500 firms across 18 variables with a known ground truth DAG. The study evaluates graph recovery metrics (F1, precision, recall) and counterfactual prediction accuracy using structural causal models. Results show consistent improvements across all algorithms when enhanced with KG and LLM knowledge. NOTEARS achieved the highest F1 score of 0.759, a 366% improvement over the baseline. The framework demonstrated perfect directional accuracy for intervention effects and a mean absolute error of 0.003610 in counterfactual predictions. Ablation studies revealed that a single focused LLM agent outperformed multi-agent setups and that algorithmic KG encoding was more effective than prompt injection. Limitations include reliance on synthetic data for evaluation and the need for temporal extensions to capture dynamic market relationships.

Research problem

Current portfolio management relies heavily on correlation-based analysis and heuristic methods, which fail to distinguish genuine causal factors from spurious correlations. Traditional statistical causal discovery methods often miss theoretically established relationships, while LLMs suffer from the 'causal parrot' problem, lacking true causal reasoning capabilities. There is a need for a hybrid approach that combines the empirical rigor of statistical methods with the domain expertise of knowledge graphs and the reasoning power of LLMs to identify true causal drivers of financial performance.

Core contributions

• A hybrid framework integrating statistical causal discovery with financial knowledge graphs and LLM reasoning.
• Algorithmic integration of KG constraints into PC, GES, and NOTEARS algorithms using edge weighting and scoring modifications.
• Demonstration of substantial performance gains: PC (+36% F1), GES (+100% F1), and NOTEARS (+366% F1).
• Reliable counterfactual prediction capabilities with MAE of 0.003610 and perfect directional accuracy for interventions.
• Insights from ablation studies showing single-agent LLM effectiveness and the superiority of algorithmic KG encoding over prompt injection.

Method

• Extraction of causal relationships from SEC 10-K filings using an agentic framework to build a financial knowledge graph.
• Formulation of KG constraints as edge weights based on strength, frequency, and coverage of causal mentions.
• Integration of KG constraints into PC algorithm via adaptive significance thresholds and protected edges.
• Augmentation of GES scoring function with KG regularization terms to bias optimization toward domain-plausible structures.
• Application of KG constraints to NOTEARS via regularization, post-hoc enforcement, and adaptive thresholding.
• Use of LLM (Qwen3-235B) 'MissingEdgeDiscoverer' to generate hypotheses for missing causal edges.
• Counterfactual estimation using linear structural equations fitted to the discovered DAG and Pearl's do-calculus.
• Ablation studies to evaluate the contribution of individual LLM reasoning modules.

Results & findings

• NOTEARS F1 score improved from 0.163 (baseline) to 0.759 with KG+LLM enhancement.
• GES F1 score improved from 0.367 to 0.735, a 100% increase.
• PC F1 score improved from 0.459 to 0.622, a 36% increase.
• Counterfactual predictions achieved a mean absolute error of 0.003610.
• Perfect directional accuracy (100%) observed for all intervention effects in scenario analysis.
• KG captured 23 of 29 ground truth edges, providing strong domain priors.
• Ablation study showed removing the 'MissingEdgeDiscoverer' module caused a statistically significant F1 drop of 0.076.
• Injecting KG information directly into LLM prompts was found less effective than algorithmic constraint encoding.

Data & experiments

• Synthetic financial dataset of 500 firms across 18 variables with a known ground truth DAG containing 29 edges.
• Evaluation of three causal discovery paradigms: PC (constraint-based), GES (score-based), and NOTEARS (continuous optimization).
• Comparison of baseline methods against KG-enhanced, LLM-enhanced, and combined KG+LLM enhanced approaches.
• Counterfactual estimation using structural causal models for six financially relevant intervention scenarios.
• Ablation studies on LLM reasoning modules to determine the impact of multi-agent vs. single-agent designs.

Limitations & caveats

• Evaluation relies on a synthetic dataset, which may not fully capture the complexity of real-world financial markets.
• The framework's performance on real-world data with noisy, incomplete, or biased 10-K filings is not tested.
• The study focuses on cross-sectional relationships, lacking temporal dynamics and lag structures.
• LLM reasoning is limited to a single 'MissingEdgeDiscoverer' module in the optimal configuration, potentially missing broader context.
• The synthetic data generation process involves assumptions about causal mechanisms that may not hold universally.
• Computational cost of running multiple LLM calls and complex optimization algorithms may limit scalability.

Curator notes

Discovery status: specific_domain_round3_promoted. Source links: FinDKG: Dynamic Knowledge Graphs with Large Language Models for Detecting Global Trends in Financial Markets