Enhancement of the predictive accuracy in logistic regression model by optimal threshold

Abstract

Diagnostic tests play a critical role in determining the default of residential mortgages. Numerous studies have attempted to screen the factors associated with the default or to examine the recovery rate of a residential mortgage. This study aims to search for an appropriate threshold probability for predicting a residential mortgage loan to be default. On one hand, some studies have used the estimated probability of 0.5 for a loan to be default; on the other hand, other studies have used the estimated probability where the lowest prediction error rate occurs. In our study, 2624 residential mortgage loans, including 249 of default and 2375 of paid-off, were collected. As for the comparison among the three thresholds, the third predictive method for binary logistic regression model provides more stable correct prediction rate, sensitivity and specificity, than the other two threshold do.

Keywords and phrases:

• Sensitivity
• specificity
• threshold
• misc/assiftcation
• total cost
• residential mortgage