Ohlson O-Score¶

Estimate a firm's near-term bankruptcy probability by feeding nine accounting and size variables into Ohlson's fitted conditional-logit model, with interpretation constrained to its population, horizon, and data definitions.

Version: v2 · 2026-09-06 · History
Domain-specific #: 2417
Origin domain: finance
Subdomain: accounting based distress modeling
Aliases: O-score, Ohlson bankruptcy score

Core Idea¶

The Ohlson O-Score is the linear predictor from James Ohlson's 1980 conditional-logit models of corporate bankruptcy. It combines nine terms derived from firm size, leverage, working capital, liquidity, profitability, operating cash flow, recent losses, and change in income. A logistic transformation maps the linear score to an estimated probability under the selected model and prediction horizon.^[1]

The construct is a historically estimated statistical model, not a timeless accounting identity. Its coefficients, scaling conventions, sample of U.S. industrial firms, reporting-era definitions, and one- or two-year horizon are part of its interpretation. A larger score generally indicates greater modeled distress risk, but calibration can drift across jurisdictions, sectors, accounting standards, and periods. Modern use therefore requires exact replication of variables and validation on the target population rather than copying a threshold from an unlabeled web calculator.

Structural Signature¶

The prediction unit. A firm-year observation is evaluated for bankruptcy within a declared horizon.
The accounting snapshot. Standardized balance-sheet, income, and funds-flow quantities supply predictors.
The size adjustment. Total assets are scaled against a price-level index in the original specification.
The solvency and liquidity ratios. Liabilities, working capital, and current-liability terms encode balance-sheet pressure.
The performance and cash-flow ratios. Net income and funds from operations encode loss absorption and cash generation.
The indicator variables. Negative net worth and consecutive-loss conditions enter as binary terms.
The income-change term. A bounded transformation represents deterioration or improvement in earnings.
The fitted coefficient vector. Historical maximum-likelihood estimation weights the nine terms.
The logistic map. The linear O-score is transformed into a model-implied probability.
The calibration envelope. Population, horizon, date, accounting definitions, and outcome labeling delimit validity.

What It Is Not¶

Not an accounting identity. Coefficients were empirically fitted and can lose calibration.
Not the Altman Z-Score. The variables, estimation method, populations, and score interpretation differ.
Not a universal bankruptcy cutoff. A threshold must reflect model version, horizon, losses, and decision costs.
Not a credit rating. It estimates a particular event probability rather than issuing an agency opinion or full credit assessment.
Not causal diagnosis. Predictor weights associate with bankruptcy risk; they do not establish why a firm will fail.
Not sufficient for an investment decision. Market value, current disclosures, liquidity, fraud, macroeconomic shifts, and decision objectives require separate analysis.

Scope of Application¶

The O-Score is a literal financial-risk instrument when its original or explicitly re-estimated logistic specification is applied to compatible firm accounting data.

Academic bankruptcy research. Benchmarking accounting-based failure prediction against later models.
Credit screening. Ranking firms for deeper review when the population has been locally validated.
Portfolio surveillance. Tracking changes in modeled distress risk across reporting periods.
Audit and going-concern analytics. Supplying one quantitative signal alongside qualitative evidence and professional judgment.
Model-risk study. Examining calibration drift across sectors, jurisdictions, and accounting regimes.
Historical replication. Reproducing Ohlson's conditional-logit design and horizons from declared variable definitions.

Clarity¶

Name the exact Ohlson model and horizon, reproduce all nine variable definitions and signs, state the price-level normalization, and distinguish the linear score from its logistic probability transform. Report how missing or denominator-edge cases are handled. Do not call an arbitrary cutoff 'the' O-Score threshold. For current use, give target-population discrimination and calibration evidence and say whether coefficients were preserved or re-estimated.

Manages Complexity¶

The score compresses a multidimensional financial statement into one reproducible risk signal, allowing large firm sets to be ranked and triaged. That compression exposes a stable audit trail from inputs to output but discards narrative context, nonlinear interactions, reporting quality, and new regimes. The concise score becomes dangerous when portability is assumed: a mathematically correct computation can still produce a poorly calibrated probability.

Abstract Reasoning¶

Fix the event definition, prediction horizon, and eligible firm population.
Map financial-statement fields to the original variable definitions and units.
Compute ratios, binary indicators, and the income-change term with declared edge-case rules.
Apply the specified coefficient vector to obtain the linear O-score.
Use the logistic transform only under the corresponding model interpretation.
Validate discrimination and calibration on the target population and period.
Select action thresholds from decision costs rather than inherited folklore.
Combine the signal with other financial, market, and qualitative evidence.

Knowledge Transfer¶

The model transfers literally only where compatible firm-level accounting fields and a bankruptcy outcome exist; even there, revalidation is required. Its broader parent is Measurement: heterogeneous indicators are combined into a scalar estimate under a declared calibration. Classification is a downstream use, not the identity of the score. Applying the formula to households, projects, or sovereigns without re-estimation is analogy or model misuse.

The score must be reconstructed from the selected Ohlson model rather than from a generic nine-variable checklist. Variable definitions include scaling choices, indicator variables, and a transformation for firm size; the coefficients belong to a particular fitted equation and forecast horizon. Changing a denominator, substituting market for book values, reversing a loss indicator, or mixing coefficients from another horizon creates a new model even if the result is still called an O-Score. Replication should preserve the equation, accounting period, units, and treatment of exceptional values.

The linear predictor and probability are different outputs. Logistic transformation maps the score monotonically into a number between zero and one, but that number is a model-based estimated probability only under the original specification and calibration. A threshold applied to the score creates a classifier whose false-positive and false-negative costs depend on use. Neither a large raw score nor a transformed probability is a bankruptcy diagnosis, credit rating, or certainty about one firm.

Transport is the central validity problem. Accounting standards, sector balance sheets, financing practices, bankruptcy law, macroeconomic regimes, and reporting quality can change predictor distributions and outcome relations. Validation on a new population should check discrimination and calibration separately. A model can rank firms reasonably while systematically overstating absolute risk, or show acceptable average calibration while failing in a sector. Re-estimation may improve local fit but produces a descendant model that should not be represented as Ohlson's unchanged 1980 specification.

Missing and delayed data also shape interpretation. Public financial statements arrive on reporting schedules, may be restated, and can reflect conditions after distress signals were already visible. A backtest must use only information available at the claimed prediction date or it introduces look-ahead bias. Excluding failed firms with incomplete records or silently imputing ratios can make performance appear stronger. The sample construction is part of the evidence ledger.

Use in decisions requires a declared action threshold and comparison model. Screening for analyst review, pricing credit, auditing going-concern risk, and portfolio exclusion carry different costs and legal contexts. The O-Score supplies one accounting-based signal, not the decision rule. Contemporary market, behavioral, or macroeconomic variables may add information, but combining them creates a broader probability-of-default model rather than expanding the identity of this node.

The parent Measurement is literal because the fitted equation constructs a scalar representation of modeled near-term distress risk from specified observables. Measurement is broader and does not fix Ohlson's variables, coefficients, sample, or horizon. Logistic Regression is the modeling family, while Classification arises only after selecting a threshold. The autonomous residual is the historically fitted nine-term instrument plus its replication, calibration, and transport boundaries.

Examples¶

Canonical¶

Ohlson estimated conditional-logit bankruptcy models on U.S. firm data and reported specifications for failure within one and two years. For a firm-year, the analyst constructs the nine declared terms, multiplies them by the published coefficients, sums the intercept and terms, and applies the logistic function when a probability is required.^[1] The resulting value is interpreted under that model's horizon and sampling frame rather than as an invariant corporate property.

Mapped back: firm-year + defined accounting inputs → nine transformed predictors → fitted linear score → logistic probability → horizon-bound distress estimate.

Applied / In Practice¶

A lender uses the published O-Score as one challenger model in a current corporate portfolio. Before routing cases, the model-risk team maps each accounting field, removes ineligible industries, measures out-of-sample discrimination, plots calibration by score band, and compares performance with a locally fitted model. The historical score remains useful if it ranks risk, but operational thresholds are set from local default costs and review capacity rather than the original sample alone.

Mapped back: current portfolio → specification-faithful calculation → local validation → cost-sensitive threshold → human review.

Structural Tensions¶

Parsimony vs. omitted context. Nine terms make the score auditable but leave out market and qualitative information. Diagnostic: What material risk signals sit outside the accounting snapshot?
Historical reproducibility vs. current calibration. Fixed coefficients enable comparison while economic and reporting regimes change. Diagnostic: When and where was calibration last tested?
Ranking power vs. probability accuracy. A model can order firms well yet misstate absolute probabilities. Diagnostic: Are discrimination and calibration reported separately?
Common cutoff vs. decision-specific action. One threshold is convenient, but optimal routing depends on false-positive and false-negative costs. Diagnostic: Which loss function justifies the cutoff?
Predictive association vs. causal explanation. Ratios forecast risk without identifying a single failure mechanism. Diagnostic: Is the score being used to triage or to claim causation?
Autonomous model vs. generic measurement. Measurement supplies the scalar-combination skeleton; Ohlson's variables, coefficients, and horizon create the named instrument. Diagnostic: Would the result still be the O-Score if those fitted commitments changed?

Structural–Framed Character¶

The O-Score is mixed. Its arithmetic is structural and reproducible, but the target event, accounting categories, firm population, sampling era, reporting rules, and action threshold are institutionally framed. It is evaluatively neutral as a model output yet often used in consequential judgments. It is human-practice-bound because bankruptcy and financial statements are legal-accounting constructs. The logistic architecture travels, while the fitted Ohlson identity remains tied to its institutional data-generating frame.

Structural Core vs. Domain Accent¶

The skeleton is heterogeneous indicators → weighted scalar → monotone probability map → thresholded decision. The domain accent consists of firm accounting definitions, bankruptcy outcome, Ohlson's nine transformations and fitted coefficients, and one- or two-year horizons. Removing those commitments yields a generic risk score or logistic model. They are precisely why the O-Score is an autonomous domain-specific abstraction rather than a prime.

Measurement is the strict parent because the O-Score constructs a scalar risk estimate from defined observables and a calibration model. Classification is related when a threshold routes firms, and Formalization describes the published equation, but neither is as taxonomically direct as Measurement.

The prospective workspace queue contains one strict upward edge to prime:measurement. No live DAG mutation is authorized.

Relationships to Other Abstractions¶

Current abstraction Ohlson O-Score Domain-specific

Parents (1) — more general patterns this builds on

Ohlson O-Score is a kind of Measurement Prime

Measurement is the strict parent because the O-Score constructs a scalar risk estimate from defined observables and a calibration model.

Hierarchy path (1) — routes to 1 parentless root

Ohlson O-Score → Measurement

Neighborhood in Abstraction Space¶

Ohlson O-Score sits in a sparse region of the domain-specific corpus (95^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Altman Z-Score. A different bankruptcy model based on discriminant analysis and a different variable set.
Probability of default model. A broader model class that may use market, behavioral, or macroeconomic predictors.
Credit rating. An ordinal agency or internal judgment rather than this fixed accounting formula.
Going-concern opinion. An auditor's professional conclusion under auditing standards, not a score output.
Logistic regression. The general model family of which Ohlson's fitted specification is one historical instance.

References¶

[1] James A. Ohlson, ‘Financial Ratios and the Probabilistic Prediction of Bankruptcy,’ Journal of Accounting Research 18, no. 1 (1980): 109–131, https://doi.org/10.2307/2490395. registry ↩a ↩b