Financial Distress Prediction Models: Key Methods Compared

A lender deciding whether to renew a credit line needs an early warning, not a guarantee that a company will fail. A risk score can flag trouble while there is still time to ask questions, adjust terms, or plan a response.

A financial ratio is a quotient that relates two figures from a company’s financial statements so we can assess its financial condition. We compare established formulas with machine learning and newer methods, focusing on decisions about U.S. companies. Some model evidence comes from other countries, so its fit for U.S. firms needs care.

We will look at what scores measure, how researchers test them, and why an impressive accuracy number may hide missed cases. The same financial risk signal can also carry different meaning across industries, time periods, and business decisions. We can use a model most effectively when we treat its result as a prompt to investigate, not as a verdict.

Key Takeaways

• A risk score flags possible trouble; it cannot guarantee a bankruptcy filing.
• Classic formulas and machine learning combine company data in different ways.
• Industry, sample, and prediction period shape what a score can tell us.
• Overall accuracy may hide missed bankruptcy or distress cases.
• Compare detection, probability reliability, and the cost of a decision.
• Use a score to guide review, not to replace it.

What Counts as an Early Warning?

A company can show strain before it misses a payment, especially when cash generation weakens while obligations remain due. The clearest early warning is a pattern across financial data, not one alarming figure by itself.

Compare at least two reporting periods to see whether a company’s financial condition is improving or weakening. Its three core financial statements are the balance sheet, income statement, and cash flow statement.

We can use each statement to check a different type of pressure. Late customer payments, upcoming bills, and weaker cash generation can tighten liquidity even when sales continue.

• Liquidity pressure: Check cash and current assets on the balance sheet against current liabilities.
• Persistent losses: Check net income on the income statement across reporting periods.
• Rising leverage: Check total debt and total assets on the balance sheet.
• Weakening cash generation: Check operating cash flow on the cash flow statement.

For a closer look at cash needs, our corporate liquidity risk guide covers obligations, cash flow, and working capital. The SEC’s Financial Statement Data Sets offer quarterly company figures filed in XBRL, from January 2009 through September 2026.

Those SEC data are presented in a flattened format for analysis and comparison, and the SEC updates them quarterly. The data may contain errors, so they do not replace a review of full filings. A single poor period can reflect a temporary disruption rather than a lasting decline.

Treat a warning as stronger when it appears across more than one statement or reporting period.

That pattern can help us separate a short-term shock from a weakening business trend. We should also note whether a large customer, seasonal cycle, or one-time cost shaped the period.

How Do Financial Ratios Turn Into Warnings?

A ratio turns two statement figures into a relationship we can compare across time or between similar companies. That comparison can reveal a change that raw dollar amounts alone may not show.

Use ratio groups to organize an initial review, especially liquidity and profitability. The same number can mean different things in separate industries, so compare like businesses and periods.

Liquidity and short-term obligations

Liquidity ratios examine whether a company can meet obligations due soon. The current ratio compares current assets with current liabilities.

If current assets rise more slowly than current liabilities, the current ratio may fall. That change can prompt us to check whether cash, receivables, or inventory caused the shift.

Current assets include items expected to turn into cash within a year, while current liabilities are due within that period. A business with slow-moving inventory may not have as much usable cash as its current ratio suggests.

We can also compare cash flow with near-term payments. A company may report enough assets yet struggle to collect receivables before bills come due.

Profitability and operating performance

Profitability ratios show how well a company earns income from its resources. Return on assets compares net income with total assets.

A falling return on assets can signal that earnings have weakened relative to the assets supporting the business. We might then inspect sales, costs, or asset use before deciding whether financial distress is growing.

Ratios are useful because they let us compare companies with different sizes. They still depend on consistent accounting periods and definitions, so a change in how a company reports an item can affect the comparison.

For a practical view of short-term cash measures, see our guide to corporate liquidity risk management.

Read a ratio alongside the business’s industry and reporting history rather than in isolation.

How Financial Distress Prediction Models Compare

The main difference is whether a method tests one indicator, combines ratios through a statistical rule, or learns a more flexible pattern from examples. Each approach trades ease of explanation against flexibility.

No model family wins for every decision; the best fit depends on the available data, audience, and purpose. Early work by Fitzpatrick appeared in 1932, while later methods joined several inputs.

Single-ratio and discriminant approaches

Univariate analysis tests one ratio at a time. Fitzpatrick’s early study helped establish this kind of comparison, which is easy to explain but can miss links among several signals.

Multivariate discriminant analysis combines inputs through a statistical rule that separates groups. Altman’s score is a well-known example, using weighted ratios to classify company risk.

These traditional financial methods are often easier to audit because their inputs and rules are visible. Their fixed structure can be limiting when relationships shift or do not follow a simple pattern.

Logistic and probit models

Logistic regression estimates the chance that a company belongs to one of two classes. James Ohlson published a logistic-regression model in 1980.

Zmijewski proposed a probit model in 1984. Probit also estimates class membership, but it uses a different statistical curve to turn combined inputs into a probability.

Both models let us examine how inputs relate to predicted risk. Their coefficients can be inspected, although the results still depend on the sample used to fit them.

What changes with machine learning

Machine learning can learn patterns from examples rather than follow only a published equation. Decision trees and support vector machines are distinct approaches, and boosting combines many smaller models.

That flexibility may help when many inputs interact in ways a fixed formula cannot capture. It can also make the prediction harder to explain to a lender, manager, or regulator.

Interpretability matters when people need to trace a decision back to a few clear factors. A flexible prediction model may suit a broad data set, but it needs careful testing for the intended use.

Start with the simplest method that matches the decision and data, then justify any added complexity.

How Does Altman’s Z-Score Work?

Altman’s original Z-score combines five ratios, and its cutoffs classify a company into distress, grey, or safe zones. The original version was designed for publicly held manufacturing firms.

A score below 1.81 falls in the distress zone; 1.81-2.99 is the grey zone, and above 2.99 is the safe zone. The original equation is Z = 1.2X1 + 1.4X2 + 3.3X3 + 0.6X4 + 1.0X5.

Altman’s original study examined 66 publicly held manufacturing firms. Half had filed for Chapter 7, and the study excluded firms with assets below $1 million.

The initial result showed 72% accuracy in predicting bankruptcy two years before the event, with a 6% Type II error rate. That result describes the original sample, not every company or period.

The original model uses market value of equity in X4. Its non-manufacturer version uses book value of equity instead, with the equation Z = 6.56X1 + 3.26X2 + 6.72X3 + 1.05X4.

In that version, X1 remains working capital/total assets, X2 is retained earnings/total assets, X3 is earnings before interest and taxes/total assets, and X4 is book value of equity/total liabilities.

Neither version should be generalized to financial companies. Their balance sheets and operating patterns differ from those of manufacturers and nonfinancial firms.

Check that a company fits the model’s intended population before treating its zone as a meaningful signal.

What Does the Ohlson O-Score Add?

The O-score uses a nine-factor weighted combination and can be expressed as an estimated probability rather than only a zone. That probability gives decision makers a different kind of output.

James Ohlson published the model in 1980, using logistic regression to combine company measures. The conversion is estimated probability = exp(O-score)/(1 + exp(O-score)).

The factors cover several parts of a company’s condition. Size and liabilities describe its scale and debt burden, while working capital and current liabilities measure liquidity.

Profitability enters through net income/assets and funds from operations/liabilities. The model also checks whether liabilities exceed assets and whether the company lost money in each of the past two years.

The last factor captures change in net income. Taken together, the inputs look at the company’s size, obligations, short-term position, earnings, and direction of change.

Because the model combines multiple measures, a shift in one input does not set the result alone. The fitted coefficients determine how much each factor contributes to the O-score.

Turning the score into a probability does not make a filing certain. An estimated probability depends on the model’s fitted coefficients and input data.

As a result, an O-score calculated from a different sample may not be directly comparable. Different company groups or periods can produce different fitted relationships.

The probability can support a lending or oversight review, but it is not a date forecast or a promise of bankruptcy. We need to keep its meaning tied to the model that produced it.

Keep the probability tied to the model and sample that produced it.

How Does the Zmijewski Model Estimate Risk?

The Zmijewski model combines three ratios, then applies a logistic probability conversion to its output. Its three inputs cover profitability, leverage, and liquidity.

A probability above 0.5 indicates a high chance of bankruptcy within one year under the model’s stated decision rule. That threshold is a classification point, not proof that a filing will occur.

The equation is Y = -4.336 – 4.513(ROA) + 5.679(FINL) + 0.004(LIQ). ROA means net income divided by total assets.

FINL is total debt divided by total assets, and LIQ is current assets divided by current liabilities. These definitions matter because reversing a numerator and denominator changes the input.

The model converts Y with P = 1/(1 + e^-Y). The resulting value ranges from zero to one, which makes it easy to compare with the 0.5 cutoff.

A value above that point marks high predicted risk under the stated rule. It does not mean the company will certainly file, and a value below it does not guarantee safety.

Unit consistency also matters. Inputs expressed with different scales can alter the result, so we should use the same ratio definitions and units throughout.

Verify each input’s numerator and denominator before interpreting a result near the 0.5 cutoff.

How Does Machine Learning Change Prediction?

Machine learning can learn patterns across many inputs instead of relying only on a fixed published equation. That flexibility can help when company signals interact in complex ways.

More flexible models still need representative examples and sound input data. Without them, a prediction model can learn patterns that fail outside its training sample.

Feature selection and nonlinear patterns

Feature selection means choosing which company measures a model should use. A review might include ratios, changes over time, and other structured financial data.

Supervised learning uses labeled examples, such as companies marked as bankruptcy or distress cases and financially healthy cases. The model learns to connect input patterns with those labels.

Unsupervised learning looks for structure without those labels. It may group companies with similar financial profiles, but those groups do not automatically mean high or low risk.

Some relationships are nonlinear: a measure may carry little warning until it crosses a certain range or combines with another factor. A decision tree can split data into such groups, while a support vector machine can draw boundaries between classes.

Those methods may help when many factors interact or when a large data set contains patterns that a fixed equation misses. The result still needs a clear test against the decision we care about.

Ensembles and model flexibility

Ensembles combine the output of several models. Boosting builds a series of models that focus on errors made earlier, which can improve classification on varied company data.

Logistic regression may be easier to audit when a lender needs to explain why a company received a certain predicted risk. Its form is less flexible, but the input effects are easier to inspect.

Complex models can be useful when the added pattern detection changes a real decision. Yet they depend on representative labels and accurate inputs, not just a large number of records.

We should also ask whether the examples match the companies being scored. A model trained on one market or period may learn signals that do not hold elsewhere.

Choose flexibility only when it improves a decision enough to justify the extra validation and explanation work.

What Can Deep Learning Detect in Company Data?

A deep neural network can learn intermediate features from financial indicators before assigning a company to a risk class. Its layers can combine simple inputs into more complex patterns.

Reported accuracy belongs to a specific sample and test design, not to every market. A 2024 conference paper tested a network on Chinese nonfinancial firms.

Layers that learn financial features

A neural network processes inputs through connected layers. Early layers can detect simple relationships among indicators, while later layers combine them into a classification.

For example, the network may learn that several financial measures together signal more risk than any one measure alone. The final layer assigns the company to a class based on that learned pattern.

Such a model can capture complex links in financial data. Its internal steps can be harder to explain than a short equation, so we need to look beyond the final class.

Reading a reported accuracy result

The 2024 study used 23 financial indicators for Shanghai and Shenzhen A-share nonfinancial companies from 2001-2018. It included 465 firms designated ST for the first time and 2,325 firms never designated ST.

The researchers used T-3 financial data to predict year T. They reported 78.33% overall accuracy and 82% accuracy for ST companies.

Those figures describe a particular market, label, lead time, and sample. The study excluded financial companies, and its results cannot be treated as a U.S. benchmark.

• Country and market: Chinese A-share firms may differ from U.S. companies.
• Sample labels: ST status defines the risk group used in this study.
• Lead time: T-3 inputs test a three-year prediction gap.
• Class balance: The sample contained far more never-ST firms than first-time ST firms.

Differences in reporting rules or company mix can change how well a network transfers to another setting. The distinction between a study’s overall accuracy and its ST accuracy also matters.

Treat the reported percentages as results for that sample and setup, not as a promised result for another market.

What Does Text Add to Financial Statements?

Narrative disclosures can add information that a balance sheet or income statement does not express in the same way. Text can describe risks, events, and management views that numbers alone may not capture.

Text and numerical data provided complementary predictive information in the ECL dataset. It combined SEC Form 10-K text, Compustat financial data, and bankruptcy labels from the LoPucki Bankruptcy Research Database.

The dataset used binary bankruptcy labels, marking each company as a bankruptcy case or not. That setup is useful for classification, but it can blur different stages of financial trouble.

In the ECL experiments, binary-label models could not distinguish filings from the year before bankruptcy from filings by financially unhealthy firms that had not yet filed.

That distinction matters when a lender needs to know whether trouble is imminent or simply present. A broad label may group both cases together, even though they call for different actions.

Text features can help when disclosures contain relevant clues that do not appear as a single financial statement item. A model might use word patterns, counts, or other text measures alongside the company’s numerical records.

In those ECL experiments, GPT-3.5 zero-shot predictions performed significantly worse than a simple keyword-based TF-IDF model. That result describes the specific test and does not establish a general verdict on language models.

A useful label should reflect the timing of the decision. If the task is to flag a near-term filing, the target should separate that outcome from long-running financial weakness when possible.

Pair text signals with numerical measures and a label that matches the timing of the decision.

Why Can Accuracy Mislead on Imbalanced Data?

A model can score well on overall accuracy by getting the most common class right while missing many of the cases that matter most. That gap is common when bankruptcy cases are rare.

Overall accuracy can stay high even when a model detects the rare class poorly. We need to know what types of errors the score counts.

Majority-class effects

A binary classifier predicts two outcomes: positive or negative. In this article, the positive class means bankruptcy or distress cases, while the negative class means financially healthy cases.

If healthy cases greatly outnumber distress cases, a model may predict the majority class often and still report strong overall accuracy. That can hide a failure to find the companies most in need of review.

We should examine a confusion matrix, which records predictions against actual outcomes. It has four cells that show correct calls and two kinds of error.

• True positive: The model predicts distress, and the company is distressed.
• False positive: The model predicts distress, but the company is financially healthy.
• True negative: The model predicts healthy, and the company is healthy.
• False negative: The model predicts healthy, but the company is distressed.

False alarms and missed cases

A false alarm can lead to extra review, tighter credit, or a lost investment opportunity. Its cost depends on the decision and the response that follows.

A false negative can leave a lender or investor unprepared for a company’s decline. The cost may be greater when a missed warning leads to losses or late action.

The preferred balance depends on what the decision maker wants to avoid. A lender protecting a credit line may accept more false alarms to catch more distress cases.

For an investor screening many firms, too many false alarms may make the process costly or impractical. We should judge the two error types in the setting where the score will be used.

Always inspect the false negatives and false positives separately before relying on a single accuracy figure.

Which Metrics Should We Use to Compare Models?

A useful model comparison tests whether the model separates risk, estimates probabilities reliably, and improves the decision being made. One metric alone cannot answer all three questions.

Compare discrimination, probability reliability, and decision utility under the same test design. A 2026 cost-sensitive bankruptcy-prediction study found that model rankings changed under decision-centered measures.

Discrimination and class detection

Discrimination measures how well a model ranks risk cases ahead of healthy cases. PR-AUC summarizes precision and recall across thresholds, which can help when positive cases are rare.

Precision asks how many predicted distress cases were truly distressed. Recall asks how many actual distress cases the model found.

PR-AUC is sensitive to the prevalence of the positive class. Comparisons across data sets with different imbalance ratios can therefore mislead.

F-score combines precision and recall into one measure. It can help compare detection, but the chosen balance may not match the real costs of false alarms and missed cases.

Calibration and decision cost

Calibration tests whether predicted probabilities match observed rates. If companies assigned a 20% risk have a much higher or lower rate in practice, those probabilities need careful review.

Expected cost estimates the consequences of decisions, including the cost of errors and actions. It connects model output to a lending or investment choice.

The 2026 study used repeated, leakage-free cross-validation and assessed PR-AUC, F-score, expected cost, and calibration measures. Boosting-based ensembles offered the most favorable balance of minority-class detection, probability reliability, and decision cost in that study.

SMOTEENN, a method that changes class balance, provided limited benefits after cost-sensitive optimization in the same study. That finding is specific to its test design and data.

Compare models under the same validation design and the same real-world cost assumptions.

How Do Economic Crises Change Model Performance?

A model can lose predictive strength when a crisis changes the relationships its coefficients learned from earlier conditions. A shock can shift company behavior, markets, and the meaning of past patterns.

Re-test the model when economic conditions change sharply. A study published March 10, 2025 examined Slovak automotive companies before and during the pandemic.

The study covered 80 Slovak automotive companies. It compared pre-pandemic years 2017-2019 with pandemic and post-pandemic years 2020-2022.

Most tested models performed worse during the pandemic. The shock changed the setting in which earlier relationships had been learned.

Researchers re-estimated coefficients, and that change improved overall accuracy and sensitivity in some models. Sensitivity measures how well a model detects the positive class.

Coefficient re-estimation can help when the input-risk relationship has shifted. It does not guarantee improvement, and the result from one sector does not transfer automatically to U.S. firms.

We can check several signs of a changed setting before carrying a model forward:

• Changed input distributions: Are company ratios or other features now outside past ranges?
• Policy or market interventions: Did new rules or support alter company behavior?
• Target meaning: Does the label still represent the same kind of risk?

A crisis may also change which companies survive long enough to appear in a data set. We should consider whether the sample still reflects the firms the model will score.

Re-test a model after a major shock before carrying its old cutoff into a new period.

What Do Recent U.S. Bankruptcy Filings Show?

U.S. business filings increased over the 12 months ending December 31, 2025, while commercial Chapter 11 filings were lower in September 2026 than in September 2025. The periods differ, so we should keep each comparison clear.

Business bankruptcy filings rose 7.1%, from 23,107 to 24,737, in the 12 months ending December 31, 2025. U.S. Chapter 11 filings totaled 9,201 in that period.

Total U.S. bankruptcy filings reached 574,314 over those same 12 months. That total includes more than business filings, so it should not be read as a count of distressed companies.

In September 2026, commercial Chapter 11 filings fell 16% year over year, from 769 in September 2025 to 648. This is a one-month comparison, not the same measure as the 12-month filing totals.

Filings show legal events, while a model estimates company-level predicted risk. Aggregate movement can provide context about the economy, but it cannot diagnose an individual company’s financial condition.

A company’s own statements, obligations, and cash generation still matter more for its specific risk assessment than a national total. A business can weaken even when aggregate filings fall.

Keep the filing category and comparison period attached to every figure.

How Can We Compare Models for a Real Decision?

We should select a model against the decision it must support, not against a leaderboard alone. The decision horizon and cost of errors shape which output is useful.

Use the same validation period and target definition for every candidate. Then test whether each model separates risk cases, gives reliable probabilities, and supports an acceptable decision cost.

A ratio score can be easier to explain when we need a quick screen. Boosting may help when structured company data contain patterns that a fixed formula misses.

A model using 10-K text and financial data may fit a review that includes narrative disclosures. It needs text from filings available by the decision date, not later information.

The prediction horizon matters. A lender deciding whether to renew credit may need a shorter warning window than an investor studying long-term portfolio risk.

Error costs matter too. A false alarm can restrict credit to a healthy business, while a missed warning can expose a lender to loss.

A strong ranking metric does not prove that predicted probabilities are useful. We must test calibration separately when decisions depend on probability values.

Reject a candidate if it cannot be validated on information that would actually have been available at decision time.

How Should We Use a Risk Score Responsibly?

A risk score supports a decision, but a person or institution remains responsible for the action taken. The score is evidence, not a final judgment about a company.

Keep the model score or probability separate from the action threshold. A score estimates risk; a cutoff determines when a decision maker takes a defined action.

A clear record helps us explain how a decision was made. We can use this short review list:

• Preserve the input date for the company information used.
• Record the selected cutoff and the action that followed.
• Review material changes in company information before acting.

Human review matters most when a score conflicts with current evidence. A recent cash loss, debt change, or major event may not yet appear in the data used by the model.

Document the score, threshold, and reason for acting so the decision can be reviewed later.

The record should make clear what the model said and what the decision maker chose. That distinction helps us spot errors and improve later reviews.

Conclusion

A useful prediction begins with a clearly defined decision, not with the most complicated model available. The model must fit the companies we score and the period we need to assess.

We should evaluate both risk detection and the reliability of probability estimates. A strong result on one measure does not settle whether a model supports the action we plan to take.

Before comparing candidates, define the decision horizon and the cost of a missed case. Write down the decision horizon and the cost of a missed warning before choosing a model to test.