Forward-looking information in GCC credit models: which variables actually explain losses

Every ECL model claims to incorporate macroeconomic forecasts. Few can show that the chosen variables explain historical losses in the portfolio being modelled.

IFRS 9 requires expected credit losses to incorporate reasonable and supportable information about forecast economic conditions. Banks, finance companies and larger corporates in the GCC have all built macroeconomic overlays to satisfy this, and regulators including SAMA expect them. The weakness in many of these overlays is that the variables were chosen by analogy with international practice rather than by evidence from the portfolio. A GDP-linked overlay borrowed from a European bank model may have little relationship to the default behaviour of a Saudi SME book.

Start from the losses

The disciplined approach is to assemble the portfolio's own history of default rates or loss rates by period, ideally quarterly over at least eight to ten years, and then to test which macroeconomic variables explain movements in that series. This requires clean data on defaults, which is itself a finding for many entities: if the default history cannot be assembled, the first investment should be in data rather than in modelling sophistication.

What tends to explain Gulf credit losses

Across the engagements we have delivered, a small number of variables carry most of the explanatory power for Saudi and Bahraini portfolios. Brent crude price, with a lag of two to four quarters, is the most consistent because of its effect on government spending and therefore on contractor and supplier receivables. Non-oil GDP growth captures the private sector activity that drives SME and retail credit. Policy interest rates, which follow the US Federal Reserve because of the currency pegs, affect debt service capacity, with a lag. For real estate-linked exposures, transaction volumes and price indices from the local authorities add explanatory power. Inflation and unemployment are frequently included but are often weakly related to losses in these markets, partly because of data quality.

The practical lesson is to test a candidate list of eight to twelve variables with their lags, keep the two or three that are statistically and economically meaningful, and document why the rest were excluded. A model with one strong, explainable variable is better than a model with six weak ones.

Scenarios and weights

Once the relationship is established, ECL is projected under multiple scenarios. Three scenarios are standard: base, downside and upside, with the base aligned to a credible external forecast such as the IMF, the central bank or a recognised forecaster, and the downside and upside constructed as plausible deviations. The probability weights applied to the scenarios are a governance decision, not an output of the model. They should be approved by the relevant committee, documented with rationale, and reviewed each period. A common weighting is 50 to 60 percent base, 25 to 30 percent downside and 15 to 20 percent upside, but the weights should reflect current conditions rather than a fixed convention.

Overlays and management judgement

Where the model cannot capture a known risk, for example a sector event or a change in regulation, a post-model adjustment may be appropriate. Regulators and auditors accept overlays when they are specific, quantified, documented and time-limited, with a plan to incorporate the risk into the model. They do not accept general prudence adjustments that persist for years. Every overlay should have an owner, a rationale and a review date.

Validation

Finally, the overlay should be back-tested. Each period, compare the ECL predicted under the scenario in effect with the losses that materialised, and record the result. Persistent over- or under-prediction is evidence the relationship or the weights need recalibration. A model that has never been back-tested has never been tested.