Forecasting under IFRS9 – Technical Challenges ahead

IFRS9 is the new accounting standard from the IASB for credit losses on portfolios of loans that is expected to come into effect in January 2018 across at least 96 of 174 jurisdictions around the globe. Work in many banks and lenders is well progressed towards meeting the reporting deadline. I will not repeat the considerations required in the building of a new provision process here as that has been well covered in many places previously. What is less well understood is the impact of the new accounting standard on forward looking regulatory submissions including various forms of periodic stress tests and forecasting exercises.

Separately there is an ongoing move internationally by regulators to enhance forecasting and stress testing. In the UK the Prudential Regulatory Authority has recently issued a consultation paper (CP17/16) that highlights the PRA’s proposed changes in the frequency and detail of forecasting to include half yearly Capital (including P&L) forecasting exercises for smaller institutions (£5bn - £50bn in assets) with a 2 year forward looking horizon. Currently these institutions are not typically required to produce this level or quality of forecast.

These regulatory pushes combined will require some mid-sized lenders to enhance the quality and governance of their forecasting procedures significantly. So what are the challenges with forecasting IFRS9? I think they fall into 3 groups:

1) Clarity of objectives
2) Definition of Stage 2
3) Incorporation of future scenarios into forecasts.

Clarity of Objectives:
One of the hardest things about forecasting future provisions under IFRS9 is that IFRS9 provisions are in themselves forward looking (ie a forecast). This makes the whole discussion quite difficult as there is a need for a precision in the language and terms used that hasn’t been required before. For this discussion I will use the terms IFRS9 “actual” calculation to refer to the forward looking provision number reported under IFRS9. I will use IFRS9 forecast of future provision to refer to a forecast of the provision fund that will be required at some point in the future under IFRS9.

Clearly it is a requirement that the economic sensitivities captured in the IFRS9 “actuals” forecast are consistent with the same economic sensitivities in the future provision forecast. Furthermore, it would seem efficient (and meet pseudo “use” tests of IFRS9) if the lifetime loss calculation is actually a component of the future provision forecast. However, a forecast of future provision requires inclusion of losses from sources not captured in the IFRS9 “actuals” number, specifically defaults from stage 1 accounts more than 12 months in the future as well losses from new business and new limit allocations. A way to isolate the contributions from these sources is a key requirement of an IFRS9 forecast of future provision.

IFRS9 actual provision calculation furthermore requires multiple scenarios to be probability weighted, this leads to a potential disconnect from a forecast of future provision where the actual scenario is a pre-condition. I will discuss this more in the consideration of future scenarios below.

Definition of Stage 2:
In the past the main complexity in forecasting and stress testing has been trying to determine the appropriate macro-economic sensitivities. This issue is no longer part of the forecasting problem however – the loss sensitivity to economic factors is already understood under the IFRS9 “actuals” calculation. Instead the key complexity under IFRS9 of forecasting future provision is the requirement to understand the timing of when losses are recognised. Specifically, when accounts move to stage 2 (defined roughly as being those accounts with significantly higher remaining life PD than anticipated at the point they were originated). Given Stage allocation is intended to be forward looking in itself and based on all reasonable evidence forecasting this is non-trivial. It will in part depend on the approach the bank has used to allocate accounts to Stage 2 – if a “triggers and rules” based approach has been used with multiple behavioural events being a criterion for stage migration then a bottom up evaluation of Stage allocation in future may be extremely challenging. Specifically, it would require all the material rules and loss parameter correlations to be captured.

For this reason, it seems rational to try to minimise the complexity of the stage allocation criteria. Notwithstanding this, Stage allocation still requires a highly granular (possibly account level) evaluation against origination expectations for every point in the forecast. Approaches for streamlining the calculation and making it computationally efficient are part of ongoing research.

Incorporation of future scenarios into forecasts:

Graph Bank UK House Price Inflation Forecasts and Data Outturns

Economic Forecast Accuracy:
Bank of England, Evaluating Forecast Performance, November 2015

Consider what we know about the accuracy of economic forecasts:

  • What we can see from the charts above is that economic forecasts have not been unbiased estimates of the future. Specifically, for unemployment at the peak of the boom in 2007 (where unemployment was lowest) the forecasts missed the uptick in unemployment. At the unemployment peak in 2009 the forecasts over estimated future unemployment. Forecasts again failed to predict the extent of the reduction in unemployment in 2013/14. On other measures the forecast process produces different types of biases – for example the bank of England have consistently under-estimated house price growth in most periods. The bank of England review finds that private forecasts perform with broadly similar accuracy but with different biases.
  • As a general rule forecasts get less accurate in recessions (when the economy is out of equilibrium) and recessions happen faster than economic models tend to predict. (See Sinclair et al 2012,
  • Furthermore, it is very likely (although difficult to prove) that models will not be self-consistent. For example, if unemployment is forecast to fall 1% each year for the next 2 years and then rise 2% in year 3, then after 2 years even if all factors turn out as expected the new economic forecast at year 2 would probably not be for 2% increase in unemployment in the 3rd
  • IFRS9 requires multiple future scenarios to be considered and a weighted average calculated within the estimation of Expected Loss. The observations above on the performance of economic forecasts suggest that simply using a single scenario, and making the assumption that forecasts are self-consistent, to estimate provision at all points in the future will be inadequate. Instead we, at least in principle, need to consider a core scenario and then at each point in time have a method for generating a range of plausible forecasts from that point on. To avoid this complexity work would be required to demonstrate that omitting this information does not lead to an under-estimate in the volatility of loan loss provisions through the economic cycle.