Loan-loss provisioning models must take a variety of economic and client factors into account, but, with the right approach, banks can develop sensible loss forecasts that are more accurate and less susceptible to volatility.
Suppose I have two competing forecasting methods, each designed for CECL or IRFS 9 loss provisioning. Both would pass muster with regulators and auditors. How do I decide which is better?
This is a critical question for builders of such systems. The statistical loss function – the criterion against which a model is judged – should always guide the modeler’s actions and inactions.
Given that CECL models seek to forecast lifetime expected losses (IFRS 9 also requires one-year expected losses to be predicted), the choice of loss function appears straightforward. The core criterion most modelers will employ will be an out-of-sample mean square forecast error or something of that ilk. The use of scenarios adds some complexity to this loss function, since the optimal forecast model must be able to cope with a wide variety of different underlying economic trajectories.
There are some further choices that must be made, like the forecast horizon over which the model will be optimized. In general, though, model validators in the CECL and IFRS 9 space will concentrate their attention on either conditional or unconditional forecast accuracy.
Before choosing a forecasting approach, there are other criteria that modelers should consider. To explore these fully, we need to step back and identify the ultimate clients of loan loss provision models.
Accounting standards are designed primarily to provide the investment community with an accurate view of the financial position and performance of the bank in question. Standing between the analysts and the investors are the senior management of the institution, whose goals are often, but not always, in line with those of investors. The third client group are regulators – and, by extension, the general public – who want the whole game to be played safely, fairly and efficiently.
All of these groups value accuracy. Investors and managers will also demand stability of the computed forecasts.
Senior executives will tear their hair out if earnings numbers jump spectacularly from quarter to quarter. They know that investors will lose confidence if they are unable to discern performance from the published accounts, causing funding costs to rise. Investors, for their part, want the financials to change if the bank experiences a material shock, but they do not want the numbers to move randomly.
By incorporating model-driven projections into the process, increased randomness is unfortunately assured. A bank with a stable portfolio facing a stable macroeconomic outlook will see earnings change every time the model is re-estimated.
Managers can either update models regularly (presenting investors with a constant drip feed of minor adjustments) or infrequently, offering quarter-to-quarter stability while risking a big jump when continued use of the model becomes untenable.
The modeler can also aim to build specifications that are less susceptible to major revision. This has been described as the holy grail of econometric model building and involves proposing specifications with exceptionally strong theoretical underpinnings. Models that result from data mining, by way of contrast, are typically much more prone to later revision.
The Recession Factor
The other source of volatility comes from economic forecasts and scenarios. Under the new accounting procedures, reserves will rise if baseline growth forecasts are marked down or if downside scenarios are accorded a higher likelihood of occurrence.
After any negative news event, you will be able to find an economist somewhere who will state that a recession is now inevitable. Sensible economists, meanwhile, move much more gingerly and will often choose not to swing at the curveballs thrown by the news cycle. Experience proves that it is very difficult to forecast a recession with any degree of confidence; it is, in fact, generally impossible to predict that a looming recession will be deep until it is already underway.
The interesting thing about the current situation is that although the U.S. economy is firing on all cylinders, a consensus is forming calling for a 2020 recession. Indeed, with tax cut stimulus timed to fade during that year, and with Congress or the Fed unlikely to provide renewed impetus, the signs are strong that a significant slowdown will soon result.
If CECL were already in place, therefore, we would expect to see relative provisions rising quite rapidly at present, with the associated curtailment of bank lending only adding to the likelihood that a 2020 – or 2019 – recession would result. CECL may improve the track record of economists in predicting recessions, but only because such forecasts will now be self-fulfilling prophecies, baked directly into the credit pie.
A 2020 recession, however, is not a foregone conclusion. In the context of the new loan loss accounting methods, and from the perspective of investors, the question to ask is whether a false positive recession prediction is better or worse than a false negative recession miss?
While this issue is clearly related to their level of risk aversion, I suspect that most investors would prefer to err on the side of stronger growth. We must remember that high capital buffers are now in effect in most jurisdictions, implying that the introduction of model-based allowances will have little cumulative impact on the safety of banks should a 2020 recession actually occur.
Because forecasts and scenarios play such a crucial role in loss provision calculation, these asymmetries are critical factors in assessing forecast error risk. CECL and IFRS 9 model managers can take positive steps to promote model stability without giving up too much in terms of accuracy. They should feed their credit loss models with economic forecasts from a sensible source that resists the notion that the sky is always falling.
If firms achieve these loss-forecasting goals, investors and senior managers should be well pleased with the results.
Tony Hughes is a managing director of economic research and credit analytics at Moody's Analytics. His work over the past 15 years has spanned the world of financial risk modeling, from corporate and retail exposures to deposits and revenues. He has also engaged in forecasting of asset prices and general macroeconomic analysis.