In a previous post, I discussed how a couple of economists at the Banco Central del Uruguay and I are identifying instances of loan evergreening—when banks provide additional credit so that firms repay their previous loans—using very granular data. The first thing we do in the paper (coming soon!) is to understand what the determinants of this strategy are.

It turns out that **the main bank-level determinant is bank solvency**. This is not entirely surprising, since a main motive to engage in this strategy is to keep provisions low. Higher provisions—which happens when a firm is delayed in its loan repayment—mean lower profits, which mean lower capital. Yet we are focusing on one particular loan evergreening strategy; there can be more, which would make it more difficult to detect the relationship with solvency. Still, we do detect it.

How do we detect it? We look at all bank-firm relations with outstanding amortizing credit at the end of each month (we focus on the period of 2006 to 2018). We then create a dummy variable that equals 100 whenever a bank is providing loan evergreening to a firm, and 0 when it is not doing so. Loan evergreening stops when the firm repays the bullet loan. Then we run a linear probability model to explain this loan evergreening variable by using bank-level variables and other controls as independent variables.

A key part of our approach is the use of fixed effects. Fixed effects regressions have become very common on papers analysing credit registers as we do. In our case, we use Firm-Month fixed effects. Let me see if I can explain this well.

Each firm, because of its characteristics, will have a probability of receiving loan evergreening. Moreover, this probability can vary across time: when the firm is performing well, maybe it is less likely to need a bullet loan to repay an existing one. In general, the situation of the firm is (or can be) an important determinant of loan evergreening. And controlling for this situation in a regression might be difficult: for instance, one could control for firm profitability, but surely this does not provide a complete picture of its situation.

Is this a big problem to understand how bank characteristics matter for loan evergreening? It might be. If banks with low solvency lend to different firms compared to banks with high solvency—which seems entirely reasonable—then it is difficult to know whether the relation between solvency and loan evergreening is due to solvency or different characteristics of the borrowers. This is where the usefulness of Firm-Month fixed effects comes in: why don’t we focus on firms borrowing from two (or more) banks at the same time? This way we make sure their characteristics are fixed, and we focus on whether banks with *less *capital are *more* likely to provide loan evergreening to the *same* firm at the *same* time compared to banks with *more* capital.

We find that banks with higher solvency (capital) are less likely to provide loan evergreening. Interestingly, the coefficient does not vary much once we add the different fixed effects: it goes from -6.578 (we only explain 0.7% of the variation) to -6.117 with all fixed effects (explaining 45.5% of the variation). These regressions include five additional bank-level controls and five loan-level controls.

There is an additional empirical check that we do to understand how robust this result is. As I just said, we have ten additional independent variables in the model other than bank solvency. We could try different combinations of these controls and see if the result stands. Or, as Brodeur and co-authors suggest, we could try them all. This is what we do. We adapt their Stata algorithm to run one regression for every single combination of these ten controls. This amounts to 2 to the power of 10 and then minus 1: **1,023 regressions**. We actually do this for all bank-level controls to see whether solvency is the most important bank characteristic to determine loan evergreening.

Ok, but then what do you do with 1,023 results? Again as suggested by Brodeur and co-authors, you can plot the resulting t-statistic—the metric that determines the statistical significance level—in a histogram and see where it lands. This is what we do in this figure:

As we can see, **solvency is the only variable whose coefficient is consistently above the 10% significance-level threshold**. Other variables are never or almost never significant. And one variable—return on assets—has around half the mass at very low values. Of course, one has to make sure that the coefficient of solvency is always negative. But again this is also easy to plot in much the same way as the t-statistics.

I mentioned before that the coefficient for solvency was -6.117. This tells you very little about the magnitude. How can we translate this in economic terms? Solvency is expressed as a number between 0 and 1: its standard deviation is 0.083 (or 8.3 percentage points). The dependent variable, loan evergreening, is either 0 or 100 (we scaled it for easier interpretation of the results). -6.117 x 0.083 = -0.508, which means that increasing bank solvency by one standard deviation implies **a reduction of the likelihood of providing loan evergreening by 0.508 percentage points**. This might not seem a lot, but actually the instances this type of loan evergreening in our sample are infrequent, just above 1%.

We find these results interesting as they provide a link—a very robust one—between solvency and loan evergreening using a very different approach to identify loan evergreening compared to the existing literature. In a future post I will talk about what happens to the firms that receive loan evergreening: are they better off, or do they end up defaulting anyway?

Pingback: Loan evergreening: does it “save” firms? | Francesc's Blog