Uncovering Why Companies Fail: An AI Approach to Bankruptcy Risk

by

Author: Denis Avetisyan

A new framework uses artificial intelligence to move beyond predicting corporate bankruptcy toward understanding the underlying causes.

ARCADIA leverages agentic AI and causal discovery techniques to analyze corporate data and identify key drivers of financial distress.

While predictive models dominate corporate risk assessment, understanding why companies fail remains a critical, yet elusive, goal. This paper introduces ARCADIA-Scalable Causal Discovery for Corporate Bankruptcy Analysis Using Agentic AI-a novel framework that integrates large language model reasoning with statistical methods to construct interpretable causal structures from financial data. ARCADIA iteratively refines potential causal graphs, yielding more reliable insights than existing algorithms like NOTEARS and DirectLiNGAM. Could this agentic approach unlock a new era of actionable intelligence in corporate finance and beyond?

Unveiling the Causal Architecture of Financial Distress

Conventional bankruptcy prediction frequently depends on statistical techniques – like logistic regression or discriminant analysis – which excel at identifying correlations within financial data but struggle to decipher the why behind a company’s decline. These methods treat financial ratios as independent variables, potentially overlooking the intricate web of causal relationships that contribute to insolvency. For example, a drop in profitability might not directly cause bankruptcy, but rather be a symptom of poor management decisions, increased competition, or broader macroeconomic factors. Consequently, relying solely on statistical correlations can lead to inaccurate predictions, especially when economic conditions shift or a firm’s circumstances diverge from historical patterns. The limitations highlight the need for models that can move beyond surface-level associations and incorporate a more nuanced understanding of the underlying drivers of financial distress, potentially through techniques like causal inference or agent-based modeling.

Conventional bankruptcy prediction models frequently encounter difficulties due to the inherent imperfections within financial datasets. These datasets are rarely complete or error-free, containing missing values, reporting inconsistencies, and, critically, data entry errors – all contributing to what is known as “noisy” data. This noise obscures genuine signals of financial distress, leading to inaccurate predictions. Furthermore, these models, often trained on historical data, demonstrate limited ability to adapt to evolving economic landscapes. Shifts in market conditions, regulatory changes, or even unforeseen global events can render previously reliable indicators ineffective, causing models to fail when faced with novel circumstances. The result is a persistent challenge in accurately identifying firms truly at risk of insolvency, even with sophisticated analytical techniques.

Predicting corporate bankruptcy demands more than simply identifying patterns in financial ratios; it necessitates discerning the causal mechanisms that propel a firm towards insolvency. Traditional statistical models frequently mistake correlation for causation, flagging symptoms rather than the root problems. A declining debt-to-equity ratio, for example, might be associated with bankruptcy, but is often a result of desperate cost-cutting measures undertaken to stave off failure. Consequently, these models struggle to anticipate distress in novel economic climates or for companies with unique characteristics. A robust predictive framework must therefore investigate the interplay of factors – managerial decisions, industry disruption, macroeconomic shocks – and how these directly influence a firm’s ability to meet its obligations, offering a far more nuanced and reliable assessment of risk than correlation-based approaches alone.

ARCADIA: An Agentic System for Causal Discovery

ARCADIA employs a Large Language Model (LLM) to generate and iteratively improve Directed Acyclic Graphs (DAGs) that model causal relationships within financial data. The LLM functions as an agent, autonomously constructing initial DAG structures representing hypothesized influences between financial variables. These graphs are not static; the LLM continually refines the DAGs based on internal reasoning and external feedback. This refinement process involves adding, removing, or modifying edges – representing causal links – and adjusting node relationships within the graph. The LLM’s capacity for iterative refinement enables ARCADIA to explore a broad search space of potential causal structures without requiring manual intervention, ultimately aiming to identify DAGs that accurately reflect underlying financial dynamics.

ARCADIA incorporates the AIDA Database, a comprehensive repository of financial statement data, to enhance the validity of discovered causal relationships. This database provides granular information on over 17,000 publicly listed U.S. companies, encompassing balance sheets, income statements, and cash flow statements dating back to 1991. By grounding its reasoning in this extensive dataset, ARCADIA moves beyond purely correlational analyses and constructs causal Directed Acyclic Graphs (DAGs) that are directly informed by observed financial behaviors. The utilization of AIDA ensures that proposed causal links between financial variables are rooted in real-world data, increasing the framework’s ability to accurately model and predict phenomena like bankruptcy risk.

ARCADIA employs an agentic loop where proposed causal graphs relating to bankruptcy risk are subjected to statistical evaluation, typically using metrics such as precision, recall, and F1-score on held-out data. The results of this evaluation serve as feedback to the Large Language Model, prompting it to revise the graph structure – adding, removing, or modifying relationships between variables. This iterative process, involving repeated cycles of graph proposal, statistical assessment, and model refinement, allows ARCADIA to progressively improve its ability to identify key drivers of bankruptcy risk and build more accurate predictive models. The framework is designed to continue this refinement process with each evaluation cycle, aiming for asymptotic convergence towards an optimal causal representation.

Statistical Validation: Ensuring Causal Rigor

The statistical evaluation within ARCADIA centers on quantitatively assessing the validity of proposed Causal Directed Acyclic Graphs (DAGs). This process incorporates multiple metrics to determine the robustness and accuracy of the discovered causal relationships. Specifically, the framework utilizes Bayesian Information Criterion differences (Delta BIC) to penalize model complexity and prevent overfitting, alongside Variance Inflation Factor (VIF) calculations to identify and address potential multicollinearity issues among the variables. These statistical measures are applied systematically during DAG construction and refinement, ensuring that the final model represents a parsimonious and statistically sound representation of the underlying causal structure.

The statistical evaluation within ARCADIA employs both Delta Bayesian Information Criterion (BIC) and Variance Inflation Factor (VIF) to assess and address potential issues in the causal DAGs. Delta BIC is used as a metric to compare the fit of different causal models, penalizing increased complexity to prevent overfitting; higher values indicate a better trade-off between fit and parsimony. Simultaneously, VIF quantifies multicollinearity among predictor variables; values exceeding a threshold (typically 5 or 10) indicate high correlation, which can destabilize coefficient estimates and hinder causal inference. ARCADIA utilizes these metrics during model construction and refinement to identify and mitigate both excessive complexity and multicollinearity, ensuring a robust and reliable causal structure.

ARCADIA consistently achieves a 100% valid-DAG rate across all tested variable budgets, indicating a high degree of reliability in its causal discovery process. This performance was evaluated by assessing the accuracy of the directed acyclic graphs (DAGs) generated by the framework against known ground truths or established causal relationships. The consistency of this result-maintained regardless of the number of variables included in the analysis-demonstrates the robustness of ARCADIA’s algorithms and its ability to effectively identify valid causal structures even in complex datasets. This near-perfect validity rate distinguishes ARCADIA from other causal discovery frameworks and confirms its capacity to produce dependable causal models.

The ARCADIA framework consistently demonstrates stable model fit, as quantified by a mean node adjusted $R^2$ value of approximately 0.10. This metric indicates the proportion of variance in each node explained by its parent nodes within the discovered Causal DAG. A consistent $R^2$ value across varying variable budgets suggests the framework’s ability to maintain predictive power even with increased model complexity. While not indicating exceptionally strong predictive capabilities in isolation, this stable performance across multiple tests supports the reliability of the identified causal relationships and the framework’s robustness to spurious correlations.

Temporal Integrity and Scalable Design

The ARCADIA framework fundamentally structures causal relationships in financial data by enforcing temporal ordering constraints within its Causal Directed Acyclic Graph (DAG). This means that a variable’s causal parents – those factors believed to influence it – must, by design, precede that variable in time. This isn’t merely a technical detail; it reflects the inherent directionality of financial events – causes demonstrably happen before their effects. By embedding this principle directly into the graph construction process, ARCADIA avoids the logical inconsistencies that can plague other methods and builds more intuitively reliable models. This approach ensures that the framework doesn’t infer, for instance, that future stock prices can influence past trading volumes, instead mirroring the realistic flow of information and decision-making within financial markets.

ARCADIA leverages the MLflow platform to meticulously document and manage every stage of the modeling process, fostering both reproducibility and accelerated iteration. This includes comprehensive tracking of parameters, metrics, and code versions associated with each experiment, creating a detailed lineage for all results. By centralizing these artifacts, researchers can easily compare different model configurations, revert to previous states, and collaboratively build upon existing work. The systematic organization facilitated by MLflow not only ensures the reliability of findings but also streamlines the development cycle, allowing for rapid prototyping and refinement of causal discovery models within the ARCADIA framework.

ARCADIA builds upon established causal discovery techniques, notably gCastle, to achieve demonstrably enhanced predictive performance and resilience in financial time series analysis. Unlike prior methods that often require substantial post-hoc pruning of identified temporal edges to remove spurious connections, ARCADIA’s architecture minimizes the need for such corrections. This capability stems from its refined Temporal Ordering constraints and robust causal graph construction, allowing it to inherently generate more coherent and accurate representations of financial relationships. The resulting framework not only improves the reliability of predictions but also streamlines the modeling process by reducing the manual intervention typically required to refine causal structures, offering a significant advancement in automated financial modeling.

The ARCADIA framework distinguishes itself through its capacity to construct remarkably coherent causal structures with minimal intervention; post-hoc pruning of disconnected nodes remains near zero, averaging only 0.350 ± 0.686 at a model complexity of M=50. This indicates a substantial improvement over existing methods, which often require extensive manual refinement to eliminate spurious connections and ensure logical consistency. By drastically reducing the need for post-processing, ARCADIA not only streamlines the model-building process but also suggests that the framework effectively captures the underlying temporal relationships within financial data, leading to more robust and interpretable causal discovery.

Runtime scales favorably with feature budget, demonstrating improved efficiency as available features increase.
Runtime scales favorably with feature budget, demonstrating improved efficiency as available features increase.

The work detailed in this paper emphasizes a move beyond mere prediction, towards understanding why a corporation might face bankruptcy. This aligns perfectly with Dijkstra’s assertion: “In computer science, the art of programming isn’t about typing, it’s about thinking.” ARCADIA, as an agentic AI framework, doesn’t simply correlate data points; it actively seeks the causal relationships driving corporate failure. By constructing Directed Acyclic Graphs (DAGs), the system maps out the underlying structure influencing outcomes – a process mirroring the careful thought required to build an elegant and robust system. Every new dependency discovered within the corporate data, as ARCADIA reveals, is indeed a hidden cost to freedom from inaccurate predictive models, and a path towards more actionable insights.

What’s Next?

The pursuit of bankruptcy prediction, now channeled through agentic AI and causal discovery, reveals a familiar pattern: a shift from seeing correlation to believing in causation. ARCADIA offers a promising step, but systems break along invisible boundaries-if one cannot clearly define the scope of causal inference within the inherently messy data of corporate finance, pain is coming. The elegance of a discovered Directed Acyclic Graph (DAG) is quickly tarnished by the realization that every node represents a simplification, every edge an assumption about a world stubbornly resisting neat categorization.

Future work must confront this inherent fragility. The focus cannot remain solely on scaling the discovery process-more data, larger models-but on rigorously defining the limits of what can be known. What unmodeled variables systematically distort the inferred causal relationships? How can one quantify the uncertainty inherent in translating natural language reports-the raw material of this analysis-into precise causal statements? The field needs to move beyond validation against prediction accuracy and embrace metrics of causal robustness.

Ultimately, the true test lies not in foretelling failure, but in understanding why systems fail. ARCADIA’s success hinges on a commitment to transparency-not just of the discovered causal graph, but of the assumptions baked into its construction. For structure dictates behavior, and a flawed understanding of that structure will inevitably lead to flawed interventions, no matter how sophisticated the algorithm.

Original article: https://arxiv.org/pdf/2512.00839.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-02 10:01