What it means
Some questions are sorting problems: healthy or failing, fraud or honest, churn or stay. Multiple discriminant analysis finds the blend of measurable variables that best separates the groups, learning from past cases whose group is known by computing the weighted combination that maximises the distance between group averages relative to their spread.
Its finance celebrity is the Altman Z-score. Edward Altman applied the method to failing and surviving manufacturers in the 1960s, and the resulting blend of five ratios still anchors credit analysis and appears in textbooks worldwide.
The output is a score and a threshold, with cases above the cut classified one way, below it the other, and grey zones between thresholds marking cases the model cannot confidently sort. The assumptions deserve respect.
The variables should be roughly normal with similar spread across groups, and the model dates: a Z-score calibrated on 1960s factories needs recalibration before judging software firms. University statistics courses teach the method as a core multivariate tool, and Penn State's applied multivariate programme devotes a full lesson to discriminant analysis, from the linear rule to its assumptions and diagnostics.
For a business owner, the value is an early-warning composite. No single ratio screams distress reliably, but a discriminant blend of liquidity, leverage, profitability and activity ratios catches patterns one ratio misses.
More than two groups are handled naturally, so a lender can sort borrowers into repayers, slow payers and defaulters in one model. Modern machine learning grew from this root.
Logistic regression and today's classifiers solve the same sorting problem with fewer assumptions, yet discriminant analysis remains valued for its transparency and interpretable weights. Validation is split-sample discipline: build the score on one set of cases and prove it on another before trusting it in the field.
That discipline separates honest models from comfortable stories, and transparency earns trust.
In practice
Real-world examples.
Example
A bank scores small-business loan applicants with a discriminant model trained on its own default history, approving borderline files only above a set score. Manual review still handles the grey zone, and the credit committee reviews the score weights each year.
Example
An auditor runs a discriminant model over engagement data to flag clients whose profile resembles past fraud cases for deeper testing. The model complements, never replaces, scepticism. Flagged files receive extra sampling rather than automatic suspicion.
Example
An investor screens a stock universe with the Z-score, shortlisting companies in the safe zone for deeper fundamental research. The screen runs monthly against fresh statements, and any company that slips into the grey zone is removed until its filings are reviewed.
Formula
Calculation
Z = 1.2 x working capital to assets + 1.4 x retained earnings to assets + 3.3 x operating earnings to assets + 0.6 x equity market value to liabilities + 1.0 x sales to assets. Scores above 2.99 read as safe, below 1.81 as distressed, with a grey zone between.
Take a struggling manufacturer with working capital to assets of 0.05, retained earnings to assets of 0.02, operating earnings to assets of 0.01, equity market value to liabilities of 0.40 and sales to assets of 0.90. Then Z = 1.2 x 0.05 + 1.4 x 0.02 + 3.3 x 0.01 + 0.6 x 0.40 + 1.0 x 0.90 = 0.060 + 0.028 + 0.033 + 0.240 + 0.900 = 1.261. That is well below 1.81, so the model places the firm in the distressed zone.Case study
Seen in the real world.
In this illustrative fictional case, Viktor, credit director of a wholesaler, screens new trade accounts with a discriminant model rather than gut feel. A fast-growing applicant passes every single-ratio check but scores deep in the distress zone on the blended calculation. He demands prepayment, and the applicant's collapse a year later costs his firm nothing. He recalibrates the model annually on his own write-off data to keep it honest.
Watch out
Common mistakes.
- Applying a model outside the population it was trained on, when a score calibrated on manufacturers misclassifies banks, utilities and young technology firms.
- Trusting the score after the inputs degrade, when creative accounting aimed at the known ratios can lift a weak firm's score without improving its health.
- Treating the grey zone as failure, when the model honestly reports uncertainty there and the right response is more investigation, not forced classification. Human judgement owns the borderline files.
Questions
People also ask.
What is multiple discriminant analysis?
A statistical method that combines several variables into a score separating cases into groups, trained on historical examples whose group membership is known. It is used for classification and prediction. The weights are chosen to maximise separation between the groups.
What is the best-known financial example?
The Altman Z-score, which blends five financial ratios to predict corporate bankruptcy. Published in the late 1960s, it remains a standard credit screening tool worldwide. Later variants adapt the approach for private and non-manufacturing firms.
What are its limitations?
It assumes specific data distributions, ages as economies change, and can be gamed once its inputs are known. Scores need recalibration for new industries and eras. Interpretability remains its strength against black-box rivals.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%Related