Corporate Credit Analysis

Module 2 worked because retail credit gives you millions of borrowers and clean outcome labels. Corporate credit inverts everything: a bank might have a few hundred corporate borrowers, each one large enough to hurt, each one different, and defaults rare enough that a statistician would call the dataset hopeless. So the discipline evolved differently — from judgment frameworks, to financial ratios, to a famous discriminant model that is still, sixty years later, the first thing anyone computes.

The Five Cs: the pre-quantitative framework

Before any model, credit officers organized their judgment around five questions:

It’s tempting to dismiss this as folk wisdom, but notice what it actually is: a decomposition of PD (character, capacity, conditions) and LGD (capital, collateral) before those terms existed. The Five Cs survive in every modern internal rating template — they’ve just been renamed “qualitative overlay.”

Ratios: which ones actually predict default

Financial ratio analysis for credit goes back at least to William Beaver’s 1966 univariate study, which asked a simple question: for each ratio, how differently is it distributed for firms that later failed versus firms that didn’t? His best single predictor wasn’t a leverage or liquidity ratio — it was cash flow to total debt. That finding has aged remarkably well.

The ratios that keep earning their place, grouped by what they measure:

• Coverage — can earnings service the debt? EBIT / interest expense is the classic; (EBITDA − capex) / interest is the honest version, because depreciation is a real cost when the assets must eventually be replaced. Coverage below 1× means the firm borrows to pay interest — that is the definition of unsustainable.
• Leverage — how big is the cushion? Total debt / EBITDA (how many years of earnings to repay), Total liabilities / Total assets (balance-sheet version). Leverage is the single most reliable default predictor across nearly every study.
• Liquidity — can they meet obligations this year? Current ratio, quick ratio. Liquidity ratios are weaker long-horizon predictors than people expect — firms rarely die slowly of illiquidity; they die when a solvency problem makes liquidity vanish overnight.
• Profitability / activity — EBIT / Total assets, asset turnover. Persistent unprofitability erodes every other cushion.

The Altman Z-score

In 1968 Edward Altman took Beaver’s one-ratio-at-a-time approach and asked the multivariate question: what combination of ratios best separates failed manufacturers from survivors? Using multiple discriminant analysis on 66 firms (33 bankrupt, 33 matched survivors), he got:

where

• $X_1$ = working capital / total assets — net short-term cushion
• $X_2$ = retained earnings / total assets — cumulative lifetime profitability; young and serially unprofitable firms score low
• $X_3$ = EBIT / total assets — current earning power, the heaviest-weighted ratio
• $X_4$ = market value of equity / total liabilities — the market’s opinion of the equity cushion
• $X_5$ = sales / total assets — asset turnover

The interpretation zones from the original paper: Z > 2.99 safe, 1.81 – 2.99 grey zone, Z < 1.81 distress.

Two things are worth appreciating. First, $X_4$ smuggles a market price into an accounting model — a hint of the structural models coming in Module 6. Second, the zones are honest: Altman published a region where the model declines to answer, which is more intellectual honesty than most modern models offer.

Z′ and Z″ — the variants you’ll actually need

The original model needs a market cap and was fitted on manufacturers. Altman re-estimated it twice:

for private firms ($X_4$ uses book equity; zones 1.23 / 2.90), and

for non-manufacturers and emerging markets — note that asset turnover ($X_5$) is dropped entirely, because turnover varies so much by industry that it swamps the signal when you leave manufacturing (zones 1.10 / 2.60). The emerging-markets adaptation adds a constant of 3.25 to map scores onto US bond-rating equivalents.

Try all three on the same balance sheet — the disagreements between them are as instructive as the scores:

Feed it a leveraged firm with strong turnover and you’ll often see Z″ flag distress while the original Z looks fine — that’s the turnover term doing the flattering. This is the calculator’s real lesson: model choice is an assumption, and reasonable models disagree on the same firm.

O-score and Zmijewski: the econometric corrections

Discriminant analysis has statistical problems (it assumes normally distributed predictors with equal covariance across groups — ratios are neither), so the next generation reached for proper probability models.

Ohlson’s O-score (1980) is a logistic regression on nine variables — size, leverage, liquidity, performance, plus two indicator variables (negative equity, recent losses). Its real contribution wasn’t the variable list; it was the output: a probability of failure, not a discriminant score. Probabilities can be priced, aggregated, and validated against realized default rates. That framing — PD as the model output — is the one that stuck, and it’s how every model since Module 2’s scorecards has worked.

Zmijewski (1984) built a three-variable probit (ROA, leverage, liquidity) but is remembered mostly for a methodological point: if you build your sample by taking all the bankrupt firms you can find plus a subset of healthy ones — which is what everyone did — your sample default rate is wildly higher than reality and your intercept is biased. Every default model built on an artificially balanced sample needs recalibration to the true base rate before its probabilities mean anything. (The same issue reappears when people oversample defaults for machine-learning models — Module 12 territory.)

The philosophical split: structural vs. reduced-form

Everything in this module treats default as a statistical pattern: find ratios that correlated with past failures, project forward. Two modules ahead — once interest rates and bonds have given us the market vocabulary — the track takes the two other possible positions:

• Structural models (Module 6) say default has a mechanism: a firm defaults when the value of its assets falls below what it owes. Model the mechanism — asset value, volatility, the debt barrier — and PD follows from economics, not curve-fitting. The price of that elegance: you need market data, and the model is only as good as its story.
• Reduced-form models (Module 7) refuse to model why firms default at all. Default is a surprise governed by an intensity, and the intensity is whatever credit-market prices say it is. Maximum agnosticism, maximum market-consistency.

Accounting models, structural models, reduced-form models: pattern, mechanism, price. Most real credit shops run all three and argue in the middle.

Where this connects

• The Z-score’s $X_4$ — market equity over liabilities — is a leverage measure priced by the market. Take that idea seriously and you get Merton’s model, where equity is an option on the firm’s assets: Module 6.
• Rating agencies formalize exactly this module’s toolkit — ratios plus qualitative overlay — into letter grades; how those grades move over time is Module 8.
• The sample-bias lesson from Zmijewski returns with teeth in the machine-learning era: Module 12.

Try it yourself

The download below is a field guide to pulling credit-relevant information out of a 10-K in about an hour: which statements to read in which order, the eight footnotes that matter most for a lender, where the bodies are usually buried (commitments & contingencies, related-party transactions, segment data), and a one-page ratio worksheet to fill in as you go. Grab a real 10-K from EDGAR and work through it once — the workflow sticks quickly.

Get new posts by email

One email per new article. No spam, no upsells, unsubscribe anytime.