Five Ratios and a Verdict: Altman's Z-Score

In the 1960s, credit analysis was a room full of people staring at ratios. Is the current ratio too low? Is the debt load heavy? Are margins thinning? Every analyst had a favourite ratio and a story about why it mattered, and every analyst reached a different conclusion.

Edward Altman, then a young academic, asked a question that seems obvious now and was radical then: instead of arguing about which ratio matters, why don't we let the data tell us which ratios matter, and how much?

The answer was the Z-score, and nearly sixty years later it is still on the first page of the credit-risk curriculum.

The problem: ratios argued, they never decided

The trouble with single-ratio analysis is that it is contradictory by construction. A company can look terrible on liquidity and excellent on profitability. Another can be swimming in cash and hemorrhaging money. Which one is closer to bankruptcy? Nothing in the toolkit told you how to weigh the good news against the bad.

Worse, nobody had ever checked whether the ratios everyone swore by actually worked. There was folklore, and there was experience, but there was no test. You could not say "companies with a current ratio below X went bankrupt Y percent of the time," because nobody had done the counting.

Altman set out to do the counting, and then to do something harder: combine the ratios into a single verdict.

The key idea via analogy: one score instead of five arguments

Think about how a doctor assesses your heart risk. They do not look only at blood pressure, or only at cholesterol, or only at age. They take several measurements, each of which is informative but none of which is decisive, and combine them into a single risk score, with the more predictive measurements getting more weight.

Altman did exactly this for companies. He took a set of firms that had actually gone bankrupt and a matched set of similar firms that had survived, and he asked a statistical procedure to find the combination of accounting ratios that best separated the two groups. The technique, discriminant analysis, is essentially a line-drawing machine: given two clouds of points, it finds the direction along which the clouds are most cleanly split, and it hands you the weights that project any new point onto that line.

The five ingredients that survived the process each capture a distinct flavour of corporate health:

1. Working capital relative to total assets. Does the company have short-term breathing room, or is it running on fumes?
2. Retained earnings relative to total assets. Has the company accumulated profits over its life, or has it never made money? This one quietly captures age and track record, because a young firm cannot have piled up much.
3. Operating earnings relative to total assets. Does the underlying business actually produce a return on the stuff it owns, before financing and tax get involved?
4. Market value of equity relative to the book value of liabilities. How far can the company's value fall before its debts exceed its worth? This is a market-based cushion, and it is the ratio that connects Altman's accounting approach to the structural credit models that came later.
5. Sales relative to total assets. Is the company generating revenue from its asset base, or is it sitting on idle capital?

Multiply each by its fitted weight, add them up, and you get a single number: the Z-score. High is healthy. Low is dangerous. In between is the famous "grey zone" where the model shrugs and tells you to do more work, which is itself an honest and underrated feature.

The remarkable part was the accuracy. Tested one year before failure, the score classified companies into bankrupt and non-bankrupt with a hit rate far higher than any single ratio, and it retained useful (though degrading) power two years out. A handful of numbers off the financial statements, combined properly, beat a room of experts arguing.

Why it mattered

• It created quantitative credit scoring. Every commercial default model, every internal bank rating system, every machine-learning bankruptcy classifier is a descendant of this idea: take observable financial data, fit weights against realised outcomes, produce a score. Altman did it with five variables and a slide rule. The method is what stuck.
• It gave the market a language. "Z-score below 1.8" became shorthand that a credit committee, a distressed-debt trader and an auditor could all understand. Shared vocabulary is how a field coordinates.
• It made expertise testable. Once you have a benchmark model, "I think this company is fine" becomes a claim you can score against a baseline. That discipline changed how credit research was done.
• It spawned a whole family. Altman himself extended it (a version for private firms without a market price, a version for non-manufacturers, a version for emerging markets), and researchers built on the approach for decades, moving from discriminant analysis to logistic regression to hazard models to gradient-boosted trees. The scaffolding is unchanged.
• It still works well enough to be dangerous to ignore. Studies over the following decades kept finding that the plain old Z-score remains a stubbornly hard baseline to beat, which is a quiet indictment of a lot of more elaborate machinery.

The honest limitations

• It was fitted on a specific world. The original sample was US manufacturing companies from the middle of the twentieth century, and the failed and healthy firms were matched by size and industry. Applying those exact coefficients to a modern software company with no factories, no inventory and an enormous intangible asset base is not what the model was built to do. The approach generalises. The numbers do not.
• Accounting is a lagging, and sometimes fictional, signal. The Z-score reads financial statements. Financial statements are published with a delay, and companies in trouble have both the motive and the latitude to make them look better than reality. Some of the most spectacular failures in history had perfectly respectable ratios right up until the fraud surfaced.
• Bankruptcy is rare, which makes the statistics tricky. In any given year, only a tiny fraction of companies fail. That imbalance makes it easy to build a model that looks accurate mostly by predicting "survives" for everyone, and it makes the model's rare-but-costly misses the ones that actually matter.
• The cutoffs are not laws of nature. The thresholds separating safe, grey and distressed came from one sample. Treating them as universal constants, rather than as calibrations that need refitting for a new era and a new industry, is the most common way people misuse the model.

The one-line takeaway

Altman showed that a handful of ordinary accounting ratios, weighted by what actually predicted failure in the data, could be collapsed into one number that flags a company heading for bankruptcy, and in doing so he invented quantitative credit scoring as a discipline.