Further evidence of regulatory distortion in Standardised and IRB frameworks

On his new blog Alt-m, George Selgin points to a piece of academic research published last year about ‘The Limits of Model-Based Regulation’, from Behn, Haselmann and Vig. This paper is very interesting and illustrates quite well how regulatory capital ratios are distorted by the use of math models encouraged by Basel 2 and 3 regulations. It nevertheless suffers from a few questionable conclusions, although those remain minor and do not affect the quality of the rest of the research.

As I described a long time ago (and also summarised in this paper), banks can calculate the risk-weighs they apply to their assets based on a few different methodologies since the introduction of Basel 2 in the years prior to the crisis. Under the ‘Standardised Method’ (which is similar to Basel 1), risk-weights are defined by regulation. Under the ‘Internal Rating Based’ method, banks can calculate their risk-weights based on internal model calculations. Under IRB, models estimate probability of default (PD), loss given default (LGD), and exposure at default (EAD). IRB is subdivided between Foundation IRB (banks only estimate PD while the two other parameters are provided by regulators) and Advanced IRB (banks use their own estimate of those three parameters). Typically, small banks use the Standardised Method, medium-sized banks F-IRB and large banks A-IRB. Basel 2 wasn’t implemented in the US before the crisis and was only progressively implemented in Europe in the few years preceding the crisis.

So what are the effects of those different regulatory capital frameworks? First, they found that

At the aggregate level, we find that reported probabilities of default (PDs) and risk-weights are significantly lower for portfolios that were already shifted to the IRB approach compared with SA portfolios still waiting for approval. In stark contrast, however, ex-post default and loss rates go in the opposite direction—actual default rates and loan losses are significantly higher in the IRB pool compared with the SA pool. […]

The loan-level analysis yields very similar insights. Even for the same firm in the same year, we find that both the reported PDs and the risk-weights are systematically lower, while the estimation errors (i.e., the difference between a dummy for actual default and the PD) are significantly higher for loans that are subject to the IRB approach vis-a-vis the SA approach. […]

Interestingly, we find that the breakdown in the relationship between risk-weights and actual loan losses is more severe the more discretion is given to the bank: while the same patterns are present for both F-IRB and A-IRB portfolios, the results are much more pronounced for loans under the A-IRB approach, which is clearly more complex and accords more autonomy to the bank.

This is pretty interesting. This demonstrates that Basel 2 (and 3) rules provide incentives to game regulatory reporting in order to maximise RoE (more on this below).

They also noticed that:

[On aggregate] to dig deeper into the mechanism, we examine the interest rate that banks charge on these loans, as interest rates give us an opportunity to assess the perceived riskiness of these loans. Interest rates in the IRB pool are significantly higher than in the SA pool, suggesting that banks were aware of the inherent riskiness of these loan portfolios, even though reported PDs and risk-weights did not reflect this. Putting it differently, while the PDs/risk-weights do a poor job of predicting defaults and losses, the interest rates seem to do a better job of measuring risk. Moreover, the results are present in every year until the end of the sample period in 2012 and are quite stable across the business cycle. […]

[Moreover, at granular loan-level] the interest rates charged on IRB loans are higher despite the reported PDs and risk-weights being lower.

This is also interesting, although I suspect partially wrong (and to be fair, they do point that out). Indeed, the German banking system is very peculiar, with small public-sector and mutual banks (likely on Standardised) having a very large market share and usually able to underprice much larger commercial banks (likely on IRB) thanks to the lack of pressure on them for profitability. Interest rates are thus likely to be understated for Standardised banks. The only way to confirm the researchers’ feeling that IRB banks charged more than Standardised ones because they knew their portfolio to be riskier is to re-run the analysis on a country with a more ‘standard’ banking system.

What is their conclusion?

All in all, our results suggest that complex, model-based regulation has failed to meet its objective of tying capital charges to actual asset risk. Counter to the stated objective of the reform, aggregate credit risk of financial institutions has increased. […]

Our results suggest that simpler rules may have their benefits, and encourage caution against the current trend towards higher complexity of financial regulation.

I cannot but only agree with this statement, despite having some doubts about the validity of their interest rate argument as explained above.

However, I will have to differ with the researchers on one particular point: that the differences we see between Standardised and IRB banks is mostly due to banks trying to game the system. Their reasoning is as follows: Basel 1 had excessively strict risk-weights, leading to ‘distortion in lending’ (absolutely agree). But the flexibility provided to banks by Basel 2’s model-based framework gets rid of this distortion and the issues described above are pretty much due to banks only (this is where I disagree).

Why do I disagree? Because of reasons I have explained before, and that are also explained within this paper: regulators validate models. Consequently, models are biased to match regulators’ expectations and biases in the first place: as it is very unlikely that regulators will consider corporate lending as less risky than real estate lending, they are also unlikely to validate a model that would do exactly this (or at least narrow the risk parameter differential). As a result, for capital optimisation purposes, banks tend to exacerbate the risk differential between those two lending types (or at least maintain the original Basel 1 risk-weight differential), in order to get regulatory approval*.

So what we’re left with is an increasingly opaque regulatory system that incentivises banks to optimise capital usage with the actual support of regulators (often against shareholders), distorting credit allocation in the meantime. Sounds effective.

*And this is exactly what the authors of this piece say!

Risk models were certified by the supervisor on a portfolio basis, and supervisors delayed the approval of each model until they felt comfortable about the reliability of the model. […]

Banks have to validate their models on an annual basis and adjust them if their estimates are inconsistent with realized default rates (see also Bundesbank 2003). Further, risk models have to be certified by the supervisor and banks have to prove that a specific model has been used for internal risk management and credit decisions for at least three years before it can be used for regulatory purposes.

PS: They also provide the following interesting chart. The same way that the introduction of RWAs triggered a real estate lending boom, at the expense of corporate lending, the introduction of Basel 2 led to a lending differential between IRB banks, which could optimise capital usage, and Standardised banks, potentially exacerbating the original real estate/corporate lending dichotomy introduced by Basel 1.

PPS: Sorry not many update recently despite having quite a lot to say… I just can’t seem to find the time to write those posts for some reason.