Everything in the nature becomes science when we understand underlying dynamics and methodology lest we think it is dark art. PPNR (pre-provision net revenue) modeling is not exception to this.
CCAR (comprehensive capital analysis and review) stress testing emerged as a regulatory requirement to be met by all BHC(bank holding company) and FBO (foreign banking organizations). As a part of this process banks need to project their losses and revenues over a nine quarter period under various stress scenarios provided by the Federal Reserve along with their own scenarios to see if bank has enough capital to meet the capital ratios. At this point, modeling revenues over 9 quarter period has emerged as hot potato and plenty of banks have received MRA/MRIA for not modeling and validating followed by review of audit in accordance to Federal Reserve standards. In this article I will provide an overview of PPNR modeling followed by some thoughts from Federal reserve guidance then describing some challenges during modeling process and how these challenges are overcome by management overlays. Finally ending with a note that PPNR modeling is not dark art but science.
Introduction PPNR:
Balance sheet of a bank provides snapshot of assets and liabilities held by the bank. On the other side, Income statement provides a measure of financial performance of a company over an accounting period. Typically a bank revenue and expense can be categorized into three areas
1) Net Interest income ( lending and borrowing activities)
2) Non –Interest income (Trading operations, brokerage activities, fund management activities)
3) Non-interest expense (Credit-Collection costs )
Under each area, there will be various individual line items for which BHC/FBO need to produce stressed loss projections.
For instance, banks need to project the growth of Consumer deposit balance in adverse and severely adverse economic scenarios. Banks need to develop quantitative models that can predict the dynamics of these consumer deposits.
Similarly banks need to project the revenues that might be made from running fixed income trading portfolio or credit trading portfolio. Below is the snapshot of industry trading revenues based of Oliver Wymann analysis across various trading segments since 1993. Banks need to project their revenues at this granularity and then aggregate to project the PPNR balances.
Banks in general build regression based models using some macro-economic variables or industry specific variables. In some cases, they use time series techniques (Auto regressive models) to predict the revenue balances. Within industry this kind of quantitative approach has become common place. Federal reserve has provided guidance around modeling practices in its August 2013 range of practices paper.
Federal Reserve Guidance on leading and lagging practice: Some important guidance provided by Federal Reserve is more of common sense rather than any special recipe.
1) Banks to have strong interaction among, central planning functions, business lines and treasury group
2) Banks have a robust challenge process
3) Banks to conduct full exploration of most relevant relationships between assumed scenario conditions and revenues and expenses.
4) Business lines expertise needs to be leveraged to build models
Overall I think modeling guide lines delineating the point that model development process does not take place in vaccum but needs lot of interaction with in the groups. For additional discussion refer to the Fed paper.
Issues in modelling : Most banks that have been conducting CCAR capital plan submissions have received MRA/NMRIA from the Federal Reserve to fix their modeling practices.
Major issues banks faced are around the limited data history and exploration of relationships between the variables and dynamics the revenue balances.
Limited data history: In general a good practice to build a model based of data that covers the economic cycle of expansion and recession. This will help in arriving in meaningful relationships instead of creating some spurious relationships or bias in the data. Because models trained using the limited data history is used in projecting 9 quarter forward looking estimates.
Selection of variables: Another important aspect of model building is selecting a set of variables that have meaningful relationship between the revenue balance and selected variables. In addition sometimes due to model developers need to perform certain statistical diagnostic tests (stationary) to check whether the selected variable is in fact suitable for modeling purposes.
Once model is developed after considering the some of these issues, model performance need to be analyzed. Due to the very nature of these models not being arbitrage free (Yield curve models) models used in pricing derivative instruments they need variety of adjustments to overall projections. These projections are in general made by business line heads.
Management overlays : In PPNR models this is very common issue that when model projections are not inline with the realized expectations then business line heads will revise these estimates. Now Federal reserve requires this kind of revision process to have clear quantitative basis and needs to be documented.
Conclusion:
So far we have understood the necessity of PPNR models. Furthermore we have also clarified the standard industry practice of using regression models and Time series techniques along with problems like data history and exploration of variable relationships. To mitigate some of the modeling issues management overlays have become common place.
To build an effective PPNR model, one needs to have good grasp of the business underlying the revenue balance.Additionally, one can use R-statistical programming package or Python programming tool to build a model. Thereofore, I think building PPNR models is more of science and less of dark art.
