Overview

The IMF study reports worse measured outcomes when rapid credit inclusion coincides with an aggregate credit boom: nonperforming loans rise, bank Z-scores fall, and GDP effects are noisier. The combined signal improves classification in the reported small sample, but AUC is not crisis probability, estimates are not cumulative forecasts, and the interaction remains predictive rather than causal.

Learning outcomes

During a concurrent boom, NPLs rise and bank Z-scores fall at the reported one- and two-year horizons.
Reported false positives were 7.5% for concurrent booms versus 11.2% for credit booms alone; AUC was 0.80 versus 0.78 in a small crisis-matched sample.
The interaction is an investigative priority, not a crisis forecast; its underlying mechanisms remain unresolved.

Keywords

interactiondistressNPLsZ-scoreAUCforecast

The Interaction That Changes the Meaning of Growth

The IMF study’s central result is not that credit inclusion is inherently destabilizing. It is that the meaning of rapid borrower growth changes when an economy is already in an aggregate credit boom. This chapter interprets that interaction in three dimensions: what happens to banking-stability measures, how much predictive information the combined signal adds, and why a useful signal still falls short of a crisis forecast.

When both booms occur, distress rises

The study estimates the effects of a credit-inclusion boom both inside and outside an aggregate credit boom. The interaction term captures the difference between those settings. If rapid borrower growth has one association when credit conditions are ordinary and another when total private credit is accelerating, the interaction is where that change appears.

The results point in the same direction across the main banking-stability measures. When a credit-inclusion boom occurs during a credit boom, nonperforming-loan rates rise by about 1.01 percentage points at the one-year horizon and 2.29 percentage points at the two-year horizon in the interaction model. At the same horizons, bank Z-scores fall by about 1.32 and 1.37 points. Because a higher NPL rate indicates more loans under stress while a lower Z-score indicates a thinner distance from default, the two measures tell a consistent story: the concurrent expansion is followed by weaker banking stability.

The magnitude matters, but so does the direction. A change of roughly two percentage points in a nonperforming-loan rate can be economically meaningful, especially when it occurs across a banking system rather than in one isolated portfolio. The paper describes the peak NPL effect as reaching about 2 percent relative to the baseline in the relevant impulse responses. That is not a prediction that every country will experience a two-point increase. It is an estimate for the study’s sample, model, horizon, and measured conditions.

The bank Z-score result adds a different kind of information. Nonperforming loans record the condition of credit already on the books. The Z-score combines a return buffer with the volatility of returns and is intended to reflect distance from default. A fall in the score does not mean that default is certain; it means that the system’s measured cushion is thinner relative to risk. Reading the two outcomes together is therefore more informative than treating either as a complete measure of stability.

The timing also changes how the coefficients should be read. The one-year NPL estimate concerns deterioration soon after the measured concurrent boom; the two-year estimate concerns a later point in the same projected path. They are not separate crisis probabilities and should not be added together. The estimates describe changes in an outcome relative to a comparison condition. A reader who turns them into a cumulative forecast would be making an additional assumption that the paper does not establish.

The signs of the measures can also confuse readers. A rise in NPLs is adverse, while a fall in the Z-score is adverse. The fact that the coefficients move in opposite numerical directions does not mean the results conflict.

Published by DeepMox

FAQ

How do NPLs and Z-scores complement each other?

NPLs show stress in loans already on the books, while the Z-score reflects a bank system’s measured distance from default; rising NPLs and falling Z-scores therefore indicate weaker stability from different angles.

What does an AUC of 0.80 mean in this study?

It means the concurrent-boom indicator had useful ranking ability in the study’s sample, not that it predicts an 80% crisis probability. The estimate comes from a small crisis-matched sample.

Why is a predictive signal not a crisis forecast?

A predictive signal associates conditions with later outcomes or ranks observations; it does not make a crisis inevitable, provide a cumulative probability, or set an intervention threshold.

Previous11 minHow Credit Inclusion Is Measured

Next11 minBeyond the Credit Warning Signal

When Credit Inclusion Meets a Credit Boom

The Interaction That Changes the Meaning of Growth

When both booms occur, distress rises

FAQ