Altman Z-Score at 50: How Credit Scoring Evolved

Corporate Financial Distress, Restructuring, and Bankruptcy by Edward I. Altman, Edith Hotchkiss, and Wei Wang (Wiley, ISBN 978-1-119-48180-5)

Previous: High-Yield Default Rates and Returns | Next: Z-Score Default Prediction and Real-World Applications

Chapter 10 is Altman’s own retrospective on the model that made his name. Fifty years after the original Z-Score paper, he walks through how corporate credit scoring evolved, why his model still gets used, and where the field is headed.

If you have ever Googled a company’s bankruptcy risk, you have probably seen a Z-Score. This chapter explains where it came from and why it refuses to die.

From Gut Feel to Multivariate Models

Credit scoring did not start with spreadsheets. In the 1800s, lenders relied on subjective factors: who owned the business, who managed it, what collateral backed the loan. By the early 1900s, rating agencies and corporate systems like DuPont’s ROE framework introduced single-ratio analysis and peer comparisons.

The real shift came in the late 1960s. Altman’s 1968 model combined five financial ratios using discriminant analysis, one of the first multivariate approaches in finance. The sample was tiny by today’s standards: 33 bankrupt manufacturers and 33 non-bankrupt ones, all with assets under $25 million. No electronic databases. Altman built the dataset by hand from annual reports and Moody’s manuals.

The five variables and their weights:

The model nailed its original sample. Firms below 1.81 went bankrupt within a year. Firms above 2.99 did not. Only three errors in the gray zone between 1.81 and 2.99.

The Family Tree Grew Fast

The original Z-Score spawned an entire family. ZETA (1977) broadened to industrials. Z’-Score (1983) swapped book equity for market equity to handle private firms. Z"-Score (1995) dropped the sales/assets ratio and added a constant, working better for non-manufacturers and emerging markets. Versions followed for SMEs, Italian mini-bonds, sovereign risk, and dozens of countries.

Parallel tracks developed too: Merton’s structural models (commercialized by KMV, later bought by Moody’s), logistic and probit regressions, neural networks, recursive partitioning, and blended models like Z-Metrics (2010). FinTech startups now layer in invoice data, payment history, and real-time signals.

Rating Agencies: Good at Birth, Slow to Update

Altman gives credit agencies their due. Original ratings are assigned carefully and create a standardized “international language of credit.” That database of ratings and transitions is invaluable for researchers.

But agencies are slow to downgrade. Studies like Altman and Rijken (2004) show that point-in-time models like Z-Score react faster to deterioration. Agencies openly prefer rating stability. Their clients (the issuers who pay for ratings) like it that way. Investors like it too, especially when the question is investment grade vs. junk.

The issuer-pays conflict of interest gets plenty of criticism. Alternative structures (investor-pays, government ratings) have not gained traction. Altman’s view: models and ratings can coexist. Models catch deterioration early. Ratings provide a benchmark the whole market understands.

Data Is King

One theme runs through the whole chapter: models are only as good as their databases. Altman built his first model from 66 firms. Today, researchers work with millions. Moody paid $3.3 billion for Bureau van Dijk’s 200 million-firm database in 2017. S&P paid $2.2 billion for SNL’s financial institution data in 2015.

The irony is that Altman’s 66-firm sample produced a model people still run on billion-dollar companies. Instructors still ask him for the original dataset. It works well enough that nobody bothered to replace it entirely.

Machine Learning: Better Accuracy, Less Trust?

Altman collaborated on a 2017 study using support vector machines, random forests, and boosting on data from 1985-2013. Adding six variables beyond the original five Z-Score inputs improved prediction accuracy by about 10%.

He remains skeptical that practitioners will embrace black-box methods for counterparty credit decisions. Complex algorithms that researchers cannot easily replicate face an adoption wall. The Z-Score wins on transparency: five ratios, fixed weights, anyone can calculate it on a spreadsheet.

My Take

Reading Altman write about his own model is like watching a parent describe a kid who left home decades ago but keeps showing up at family dinners. He is proud. He is also honest about the limitations. The 1.81 cutoff made sense in 1966. It misclassifies a lot of firms today.

What strikes me most is the durability. Hundreds of academic papers have built “better” models. Machine learning beats discriminant analysis on test samples. Yet Bloomberg and S&P Capital IQ still get more equity analyst hits on the Z-Score page than bond analyst hits. A model built on 66 small manufacturers in the 1960s still drives investment decisions in 2018.

The next post covers what happened when Altman tried to move beyond “safe vs. distress” zones into actual probability-of-default estimates, and how the model got used in the real world (including his testimony before Congress on General Motors).

Previous: High-Yield Default Rates and Returns | Next: Z-Score Default Prediction and Real-World Applications