How questionable is the comparability of Basel risk weights in the EU banking sector?

In a couple of weeks’ time, at the FEBS conference in Rome, Stefan Kerbl and Zsofia Döme from the Austrian national bank (OENB), will present a paper on the comparability of Basel risk weights in the EU banking sector. The short story: according to Stefan and Zsofia there are significant differences in the ways European banks risk-weigh their assets. So much so, that both authors question the extent to which EU bank rules harmonize bank practices in Europe. The authors then claim that the country where a bank is headquartered creates statistically significant and economically important differences. These are interesting findings indeed!

A graph that features prominently in the OENB study is the one below:

Kerbl graph 1

The Standard Approach, in blue, has distinctive concentrations around 75 and 100 percent; indicating lending to SMEs and unrated corporates. In red, the IRB risk weights are significantly lower, which then frees up expensive bank capital.

Not inconsequential

The graph confirms the narrative that banks cannot really be trusted to handle the weighing of risks on their own. Previous studies have confirmed this. They led the Basel committee to limit the use of internal models. The ECB responded to studies like these with an initiative to trim the variability in risk weights. In other words: such studies are not inconsequential.

The graph then prompts the authors to dig a bit deeper. They decide to examine the way banks apply the methods for risk-weighing assets per member state.

Time for a replication

I found the graph odd, because the differences of these average risk weights are so … stark. Just imagine how risky a bank would be if these averages apply to all assets of the bank. There is also that large gap between SA portfolios at 100% and the associated data point for IRB portfolios. If this is lending to corporates in OECD countries, which may represent a large volume of lending, then the graph tells us that IRB banks apply far lower risk weights to this category of lending than SA banks do. Moreover, the graph seems to imply that SA banks leave money on the table by not using the IRB approach. The SA banks could free up capital by switching to IRB.

What is going on?

With access to the same EBA data as Stefan and Zsofia, I decided to replicate their study. Credit to the authors, their explanation of the research design is clear.

Below is my version of the graph, where I superimposed histograms of the portfolio risk weights. As you can see, the kernel densities are broadly the same: the average IRB risk weights are low; the SA risk weights are high.


The differences between my graph and the one of Stefan and Zsofia are the result of the kernel density algorithm settings and the choice of granularity of the portfolio data – for example do the authors look at Corporates in aggregate or at the specific portfolio labeled “Corporates”? That was not clear to me. Still, my replicated density plots confirm the story of Stefan and Zsofia.

Raw averages indeed

However, one thing that I found weird in the OENB study is that, for the graph, the authors do not value-weigh the portfolios. The graph shows the raw, unweighted, average values of all portfolios. Whether a bank invests one euro in a portfolio or one billion euros, it does not matter. Uh, oh. A bank with many small portfolios that benefit from the IRB treatment will move the averages to the left of the graph, even though the total amount invested in these portfolios, in proportion to all funds invested in portfolios, is relatively small. This is weird.

It would therefore be interesting to see what happens if one adjusts each portfolio for the proportion of that portfolio to all other portfolios. This should give a more representative picture: tiny portfolios are not able to bias the results. Here is the result based on value-weighted portfolios:

value-weighted averages

Now the results are a much more harmonious! There is basically no difference between banks using either the IRB or the SA approach. The kernel density graphs are almost indistinguishable. More importantly, there is no pushing down of risk weights by IRB banks anymore. In fact, the SA portfolios seem to be the ones that are pushed down (according to the histogram). But again, the differences are minimal if one looks at the kernel densities. These harmonious results are likely the result of banks choosing their models more carefully: IRB banks may control their exposures in such a way that they fit their business model, or the other way around. However, examining endogeneity issues is beyond the scope of this post.

Given that a replication using value-weighted portfolios shows that EU is more harmonized than Stefan and Zsofia argue, I am therefore not so sure what both authors are up to.