In 78 percent of the cases, the limit granted is smaller than the amount permitted. Most strik- ingly, in 64 percent of the cases for which we know the amount granted in the previous period, the amount granted is exactly equal to the amount granted in the previous period (it is smaller 4 percent of the times, and goes up only in 34 percent of the cases). Given that that inflation rate was 5 percent or higher, the real amount of the loans therefore decreases between two ad- jacent years in a majority of the cases and to make matters worse, in 73 percent of these cases the firm’s sales had increased, implying, one presumes, a greater demand for working capital. Further, this is the case despite the fact that according to the bank’s own rules, the limit could have gone up in 64 percent of the cases (note that getting a higher limit is simply an option and does not cost the firm anything unless it uses the money). Finally, this tendency seems to become more pronounced over time: in 1997, the limit was equal to the previous granted limit 53 percent of the time. In 1999, it did not change in 70 percent of the cases.

[TABLE 7 ABOUT HERE] In Table 7, we regress the actual limit granted on information that might be expected to play a role in its determination. Not surprisingly, given everything we have said, past loan is a very powerful predictor of today’s loan. The R-squared of the regressions is also very high (over 95 percent). In column (1), we regress (log) current loan amount on (log) past loan amount and the (log) limit according to the rules. Note that the bank’s rule never refers to past loan as a determinant of the loan amount to be given out. Yet the coefficient of past loan is 0.757, with a t-statistic of 18 (a one percent increase in past loan is associated with a 0.756 percent increase in current loan, after controlling for the official rule). The maximum limit is also a significant determinant of loan amount, with a coefficient of 0.256. The standard deviation of these two variables is very close (1.50 and 1.499 respectively). These coefficients thus mean that a one standard deviation increase in the log of the previous granted limit increases the log of the granted limit by 3 times as much as a one standard deviation increase in the log of the maximum limit as calculated by the bank.

In column (2), we “unpack” the official limit: we include separately the bank’s limit on turnover basis (LTB), the limit based on the traditional method (LWC), and now include the logarithm of profits. As in the previous regression, past loan is the most powerful predictor of current loan. Both limits enter the regression. Neither the log of profit nor the dummy for

23