Betting On A Credit Bubble? Your Portfolio Might Bust First.

| About: Financial Select (XLF)
This article is now exclusive for PRO subscribers.


Bank portfolios are robust than most analysts' believe.

Late payments are not linked to one macro variable and correlations fall to zero in extreme markets.

A perfect storm of economic headwinds is required to bust this bubble.


This article examines the relationship between bank loan delinquencies and bank charge-offs vis-à-vis credit spreads and other macroeconomic variables over 30+ years of data. Some indicators like credit spreads do not explain extreme delinquencies rates. If you bet on a credit bubble, you are going to need more than one silver bullet.


Since the Fed data for bank charge-offs and loan delinquencies has been available since 1985, I attempted to find data over the same period for the other independent variables. To test my hypotheses, I relied on two forms of analysis. First, I plotted charge-offs and loan delinquencies against each of the macroeconomic factors to develop a view on how changes in each variable track changes in the other. Second, I regressed each indicator on the others to determine the statistical and predictive strength of each indicator.

Bank Loan Delinquencies vs. Credit Spreads

Credit spreads are defined as the spread between Moody's Baa 10 Year Yield over 10-year U.S. Treasury Notes. Figure 1 shows that bank loan delinquencies generally have a positive relationship with credit spreads. Macro credit risk is reflected in these spreads - widening spreads indicate growing concern about the ability of corporate borrowers to service debt, while narrowing spreads indicate falling credit risk. Since credit spreads may predict future economic activity, it is not a surprise that high bank loan delinquencies follow increases in credit spreads (for example, the widening of credit spreads in late 2007 is followed by a spike in bank loan delinquencies). As credit spreads fell from around 6% in late 2008 to around 3% by mid-2009, bank loan delinquencies also fell dramatically.

Figure 1

Source: Duke academic data, Federal Reserve databases

Figure 2

Source: Duke academic data, Federal Reserve databases

Theory suggests that widening credit spreads should result in an increase in delinquencies. However, in my regression of bank loan delinquencies to credit spreads, the R-squared is only 0.04, though the coefficient of 0.45 (2.33) supported my beliefs (more on this later). Nonetheless, a visual inspection of the data suggests that credit spreads might have more explanatory power at lower delinquency rates. Consequently, I split data into two sets: periods with delinquency rates below 3% and those above 3% (the average over the entire period was 3.66%).

Figure 3

Source: Duke academic data, Federal Reserve databases

It appears that credit spreads explain delinquency rates at low levels. However, credit spreads are a poorer explanatory factor of delinquency rates above 3%. Therefore, credit spreads over a certain level overstate the risk of delinquencies, mitigating its accuracy as an indicator in extreme environments.

Bank Charge-offs vs. Credit Spreads

Similar to bank loan delinquencies and credit spreads, bank loan charge-offs and credit spreads are closely related. For example, in late 2007, both variables spiked.

Figure 4

Source: Duke academic data, Federal Reserve databases

Figure 5

Source: Duke academic data, Federal Reserve databases

Once again, I observed a decoupling between spreads and charge-offs, in periods when charge-off rates were relatively high. Therefore, I separated the data into two sub-groups, over and under a level of 1%.

Figure 6

Source: Duke Academic data, Federal Reserve databases

Credit spreads alone cannot account for high levels of missed payments on loans, which appears to support my findings with delinquencies. My regression model predicts that the charge-offs (in normal environments) would be near zero in the absence of credit spreads. One interpretation of the coefficient value is that almost all firms rated at Baa are capable of generating cash at the risk-free rate in normal business environments. Another is that firms have structured their net interest payments so that they are immune to increases in the risk free rate. If the latter was correct, then I would expect the coefficient to be zero in the comparable regression on delinquency rates - I did not. One could also attribute the difference in coefficients to a measure of hedging ineffectiveness.

Bank Loan Delinquencies vs. Bank Stock Returns

As bank loan delinquencies rise, I expect bank stock returns to decline. While this relationship would at first seem straightforward, Figure 7 clearly demonstrates otherwise (the available KBW data were only available through 2011).

Figure 7

Source: Duke Academic data, Federal Reserve databases, KBW

Figure 8

Source: Duke Academic data, Federal Reserve databases, KBW

One possible explanation is that stock returns are simply too noisy. Another is that since stocks are a leading indicator of performance because they reflect forward earnings, thus the relationship will not contemporaneous. In the name of thoroughness, I opted to test the both, using six-month trailing returns as well as average returns over the year.

Figure 9

Source: Duke Academic data, Federal Reserve databases, KBW

Unfortunately, neither regression provided encouraging results. One cynical interpretation of the results is that if stock prices reflect analysts' estimates, then the estimates are poor predictors of potential losses. I prefer to interpret the results as a reflection of the noise in stock prices. The third regression in Figure 9 regressed delinquency rates below 3% versus bank returns and was included for completeness. While some permutation of the three might provide interesting results, I opted against data mining.

Bank Charge-offs vs. Bank Stock Returns

As Figure 10 illustrates, charge-offs are also immune to the volatility of stock returns.

Figure 10

Source: Duke Academic data, Federal Reserve databases, KBW

Figure 11

Source: Duke Academic data, Federal Reserve databases, KBW

I opted against conducting additional regressions on transformed bank returns and charge-offs. Since banks are unlikely to record charge-offs against performing loans, if bank returns were not a good predictor of delinquencies, then they are unlikely to be a good predictor of charge-offs.

Bank Loan Delinquencies vs. Unemployment Rate

Unemployment and loan delinquencies display a consistent positive relationship. The graph in Figure 11 suggests that the unemployment rate rises almost in tandem with bank loan delinquencies. Naturally, as the unemployment rate rises and more people lose their jobs, their ability to meet debt payments will decrease. In the three recessions in my data (July 1990 - March 1991, March 2001 - November 2001, and December 2007 - June 2009), an uptick in the unemployment rate is quickly followed by a rise in loan delinquencies.

Figure 12

Source: Duke Academic data, Federal Reserve databases

Figure 13

Source: Duke Academic data, Federal Reserve databases

Bank Charge-offs vs. Unemployment Rate

I would expect bank charge-offs to be positively correlated to the unemployment rate, and the data since 1985 confirms my hypothesis.

Figure 14

Source: Duke Academic data, Federal Reserve databases

Figure 15

Source: Duke Academic data, Federal Reserve databases

Bank Loan Delinquencies vs. Household Net Worth

Intuitively, I would expect bank loan delinquencies and household net worth to be negatively correlated. As household net worth decreases, bank loan delinquency would likely increase, holding aside differences in the timing of statistical reporting of these indicators. While Figure 16 illustrates this relationship at times (see the Great Financial Crisis).

Figure 16

Source: Duke Academic data, Federal Reserve databases

Figure 17

Source: Duke Academic data, Federal Reserve databases

Regressing delinquencies against contemporaneous net worth did not provide encouraging results. Consequently, I used twelve-month trailing net worth. Although R-squared came in at the lower end of my expectations, it did increase from near zero (in the contemporaneous regression). Furthermore, the t-stat of the coefficient also moved into statistically significant territory (2.75). Given home prices constitute a large portion of consumers' net worth, I believe that the results are largely reflective of liquidity constraints.

Bank Charge-offs vs. Household Net Worth

I would expect bank charge-offs and household net worth data to be negatively correlated, as with delinquencies. However, I hope to see a stronger relationship than with delinquencies since borrowers would have more time to overcome illiquidity constraints.

Figure 18

Source: Duke Academic data, Federal Reserve databases

Figure 19

Source: Duke Academic data, Federal Reserve databases

Running a regression yields an R-squared of only 0.20, though the t-statistic is -5.35. It should be noted that using contemporaneous net worth yielded poorer results than using the four quarter moving average. I transformed the net worth data, based on the same reasoning as with delinquency rates.

Bank Loan Delinquencies vs. Housing Prices

Despite illiquidity problem described above, Figure 20 below does a better job of reflecting this inverse relationship (as falling home prices, measured by the S&P/Case Schiller Index, portends a rise in bank loan delinquencies, particularly in late 2007).

Figure 20

Source: Duke Academic data, Federal Reserve databases

Figure 21

Source: Duke Academic data, Federal Reserve databases

The results from my regression on delinquencies were somewhat positive (coefficient of the t-stat was -3.6). The coefficient of home prices has one interesting interpretation - it's the rate at which an increase in home prices could be converted to cash.

Bank Charge-offs vs. Housing Prices

The data depicts a similar picture as with delinquencies.

Figure 22

Source: Duke Academic data, Federal Reserve databases

Figure 23

Source: Duke Academic data, Federal Reserve databases

Bank Loan Delinquencies vs. GDP

Loan delinquency and GDP should be negatively correlated. As GDP rises, people should be better able to repay their loans and delinquencies should fall. It should be noted that I implicitly assume that labor's share of GDP (70%) remains constant over time. Whilst recent data indicates that this might be a strong assumption going forward.

Figure 24

Source: Duke Academic data, Federal Reserve databases

Figure 25

Source: Duke Academic data, Federal Reserve databases

The regression results depicted in Figure 25 implies that there is a weak relationship between delinquencies and real GDP growth. I opted against segregating the data by delinquency rates (as I did above) because the distribution of delinquency rates for a given GDP growth rate range is equally distributed above and below the mean.

Bank Charge-offs vs. GDP

The bar chart below in Figure 26 clearly shows the negative correlation outlined above.

Figure 26

Source: Duke Academic data, Federal Reserve databases

Figure 27

Source: Duke Academic data, Federal Reserve databases

My results for charge-offs and GDP growth are in line with my results for delinquency rates vs. GDP growth. However, the GDP coefficient in this instance was significant coming in at -2.76.

Bank Loan Delinquencies vs. Consumption

Here I presumed the relationship assumed to be negatively correlated, high delinquency rates should be associated with falling consumption levels. The reasoning is simple; as the economy improves and consumers' income increases, their ability to meet their financial obligations also improves in tandem. Hence there is a reduction in banks' bad debts and corresponding charge-off amounts.

Figure 28

Source: Duke Academic data, Federal Reserve databases

Figure 29

Source: Duke Academic data, Federal Reserve databases

A cursory examination of the data revealed that quarterly change in consumption is extremely noisy. Consequently, to smooth the results I opted to use the average change in the last four quarters. Consequently, t-stat moved into significant territory and the R-squared improved, as illustrated by Figure 29. Similar to GDP growth, delinquency rates are almost equally distributed above and below the mean for a given level of consumption growth.

Bank Charge-offs vs. Consumption

As shown in the graph below, bank charge-offs are negatively correlated to consumption.

Figure 30

Source: Duke Academic data, Federal Reserve databases

Figure 31

Source: Duke Academic data, Federal Reserve databases

When I initially ran a regression between consumption and charge-offs, the results were poor; the R-squared was low and the t-stat was insignificant. When I examined the relationship more thoroughly, I realized that the consumption data from quarter to quarter was highly volatile. Thus, I opted to take the average change over the last four months to reduce noise. The regression with the smoothed data seemed to conform more to expectations: R-squared increased from 0.019 to 0.33. Furthermore, the t-stat for consumption was significant (-7.6).

Multi-factor regression on delinquency rates

After estimating the single factor regressions, I constructed a multi-factor regression using the variables I believed had the most explanatory potential versus delinquencies. Namely, I picked average consumption change over the last four quarters, the unemployment rate, and household net worth change over the last four quarters.

The summary statistics for the regression are presented in Figure 32. The overall regression had an R-squared of 0.67 and all the variables' t-stats were significant at the 10% level. Surprisingly, the coefficient of consumption was positive. While this seems to run counter to economic theory, I believed that there were two economic explanations.


Source: Duke Academic data, Federal Reserve databases

One of my hypotheses was that banks loosen credit standards in good economic times; a sustained increase in consumption makes banks less risk averse. My second hypothesis was based on a belief that banks see consumers and businesses as two distinct credit risks. Consequently, lower implied risk to one group meant that banks could take more risks with the other. I believe that the second hypothesis was less likely to be true, since my single factor regressions indicate consumption and delinquency rates are negatively correlated. Thus, my second hypothesis could only be true in the event that the unemployment rate and household net worth explained the lion's share of the perceived risk of lending to consumers. Note that I eliminated multicollinearity as a possible explanation.

Next, I constructed a test that could test both of my hypotheses. I estimated a regression using the same dependent variables against delinquency rates for consumer loans. If my first hypothesis was correct then I should observe a positive coefficient for consumption in both regressions. If, however, the coefficient was negative, then it would appear my second hypothesis was correct.

The results of the regression appear in Figure 33. Surprisingly, the results of the new regression supported my second hypothesis. Thus, it appears that banks try and maximize profits by making riskier bets with businesses when the risk of lending to consumers falls.

Figure 33

Source: Duke Academic data, Federal Reserve databases

Multi-factor regression on charge-off rates

Once again I wanted to construct a multi-factor regression for charge-offs similar to delinquency rates. However, since banks have discretion over charge-offs (accounting standards allow for management to use their judgement to estimate charge-offs), I need a proxy for an impetus for a bank to record a charge-off.

Unfortunately, bank returns, which theoretically could have served as a proxy, did not live up to its promise. As discussed above, I suspect that the main reason for the statistically insignificant results (see loan delinquencies vs. bank returns) was primarily due to the volatility of bank stock prices. To reduce the noise associated with the returns data, I constructed a dummy variable called "sustained losses". The decision rule was as follows: if average bank returns for the last four quarters was negative, then I could categorize it as a period of sustained losses and vice-versa. I hypothesized that in periods of sustained losses banks would be more likely to record charge-offs for the following reasons;

1) Sustained losses would result in more conservative assessments of their accounts.

2) In periods of sustained losses, a charge-off has a lower relative impact on a firm's already depressed stock price.

I added two other variables to the regression: twelve-month trailing unemployment and the four quarter moving average of consumption. The results of the regression are listed below.

Figure 34

Source: Duke Academic data, Federal Reserve databases

Unfortunately, the sustained losses variable was not significant at the 5% level. However, I chose to include the variable in my final model since it is theoretically warranted, and it is significant at the 10% level. The regression appears to support my intuition: banks are more likely to record charge-offs in periods of sustained losses and high levels of unemployment. Furthermore, reductions in consumption also lead to increased charge-offs.

My results indicate that the economic drivers of charge-offs occurred in the distant past and that contemporaneous economic activity might be disconnected from charge-offs. Consequently, banks may record losses as a result of a downturn even after the economy was returned to growth.


While my graphical and statistical analyses yielded mixed results in light of well-accepted economic theory, those that did not confirm my hypotheses told important stories about the lagged impact of certain variables. Specifically, delinquency is more contemporaneous linked to economic activity, and charge-offs trail by 12 months (this is confirmed by the higher R-squares when I regressed 12-month trailing). More importantly it highlighted that some variables lose some of their explanatory power as we transition from one economic state to another. Therefore, an analyst should view unconditional explanations of bank profitability with a healthy degree of caution. It is likely that a perfect storm of macro headwinds will be required to bust the credit bubble, as opposed to one silver bullet.

Disclosure: I/we have no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours.

I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.

Additional disclosure: Special thanks to Alvin O., Josh A, and Raja S. for helping gather the data and creating the charts.