Tobit regression, also known as the censored regression model, is a technique used when the observed dependent variable is censored, meaning that the observed values are restricted above or below a certain threshold, but the unobserved latent dependent variable can take any value. The model estimates a linear relationship between the unobserved latent variable and the independent variables, taking into account the presence of censored data points.
A practical application of the Tobit model is to predict Loss Given Default (LGD) by accounting for the limits of the LGD variable, typically 0% to 100%. In a default scenario, the Tobit model predicts the expected loss for a loan, where the response variable (LGD) is bounded by total recovery (0) and total loss (1), making it suitable for modeling the partially observed nature of recovery.
- tobit.py: implementation of Tobit model, informaed and improved upon the implementation by James Jensen.
- tobit_model_lgd.ipynb: a notebook that provides an example of using Tobit Model (censored) regression to Loss Given Default (LGD) by accounting for the limits of the LGD variable.
- Tobit_Model_Notes.pdf: a research note that reviews the assumption, formulation, parameter estimation, and prediction of the Tobit model for censored regression.
The project is inspired and informed by the following sources: