Core Path · L6.3.02

Logistic regression for default

14 minLogistic regressionMaximum likelihoodOdds ratioAIC and BICStepwise selectionCollinearity

Know these first

By the end you will be able to

  1. 1Explain how the logit link turns a weighted sum of inputs into a PD, and read a coefficient as a change in log-odds and an odds ratio.
  2. 2Fit a logistic regression by maximum likelihood and read its standard errors, Wald and likelihood-ratio tests, AIC and BIC.
  3. 3Choose inputs with stepwise selection and credit judgement, and spot collinearity from VIFs and flipped signs.
  4. 4Compare fitted weights with expert weights, separating how well a model ranks from how well it is calibrated.

Why it matters

RiskAtoms Bank plc's corporate rating model (L3.3.04: Internal ratings and master scales) puts a weight of 0.55 on net leverage because its designers judged it so. Fit the same four inputs to 22,266 firm-years of history and the data say 0.254. The judged weights rank firms well, but on this panel they predict an average PD of 7.6% where 2.55% defaulted. L3.3.02: Analysing a company's credit promised that statistical models would turn ratios into probabilities. This lesson shows how, and what the fitted weights can and cannot tell the bank.