Choosing Long Macro Lags in IFRS 9 Point-in-Time PD Models
Summary
The document considers whether a point-in-time probability-of-default model can include macroeconomic variables whose effects arrive more than a year later. Its example models next quarter’s default rate using the current default rate and a lagged change in GDP. With a lag of five quarters, GDP forecasts for the coming year would affect default rates beyond that year, leaving the near-term estimate dependent on observed history rather than the new forecasts. This creates difficulty when producing scenario-specific estimates or stress tests intended to reflect a recent shock.
The response says long lags can still belong in a point-in-time model, particularly when the goal is to capture persistent macroeconomic effects or changes in expected credit losses. It suggests time-series regression, generalized additive models, or recursive models as ways to examine seasonality, trends, and temporal confounders. If the task is to measure the immediate effect of a recent change, shorter lags are more suitable. The exchange gives modeling guidance, but no fitted results or validation evidence, and the right lag structure depends on the intended use.
Key ideas
- A long macroeconomic lag can make a next-year PD estimate insensitive to forecasts for the coming year.
- Lagged macro variables may still help capture persistent effects in point-in-time PD estimates.
- Shorter lags are more appropriate when the goal is to reflect recent economic changes promptly.
- Time-series regression, generalized additive models, and recursive models can help examine lag structure and temporal patterns.
- Model selection should reflect whether the objective is near-term scenario sensitivity or longer-term effects on expected credit losses.
Tags
Full text
# IFRS 9 PiT-PD model when the lagged dependency exceeds one year
# IFRS 9 PiT-PD model when the lagged dependency exceeds one year
It is a common approach to model the point-in-time PD (PiT PD; meaning that the PD depends on the current or lagged economy) by regressing default rates on current or past macro variables (such as GDP or unemployment rate). The specific functional form (e.g., regression of logits) is not relevant for my question. Let us assume that we found a model of the following form: $$ DR_{t+1} = f(DR_t, \Delta GDP_{t-5}) $$ where $t$ measures quarters. This means that the default rate of the following year depends on the current default rate (a sensible assumption) and the change of GDP five quarters ago.
In order to incorporate forward looking information into our PiT-PD estimate of the following year we consider $\Delta GDP$ forecasts for $t+1, \ldots, t+4$. Looking at the above equation and due to the lag in the reaction, these forcasts have, according to our model, impact on $$ DR_{t+7}, DR_{t+8}, DR_{t+9} \text{ and } DR_{t+10}. $$ Finally, this means that the whole PiT-PD for the coming year is indpendent of the forecast and is rather a deterministic calculation using observed GDP changes from $t-5$ until $t-1$.
While this looks simple, we can not use this for the following use cases for the coming year:
- incorporating various forward looking forecasts (scenarios) to get different PDs in these scenarios.
- Stresstesting by assuming that a future GDP decrease impacts next years default rate in our portolio.
My question, thus, is:
- Can we use PiT-models with lagged relations of more than one year at all for PiT-modelling. It looks as this then only helps for the lifetime view.
- Should we restrict feasible models to such where a timely reaction is assumed?
Happy to read any comments on this.
## Answer by Jean Dessain (score 2, accepted)
https://quant.stackexchange.com/a/77618
I will try to provide an answer, even if it should most likely be refined with the detailed case you are working on:
- It is reasonnable to use PiT models with lagged relations of more than one-year. To predict the PiT PD, you can perfectly integrate lagged events older than a year. However, as you identified, the effectiveness of such model is stronger for TTC PDs as it then fully captures the long term impact of changes in the macro-economic environment. Performing time-series regression, GAMs analysis or even recursive models, is a way to model the seasonality and LT trends and identify temporal confounders for investigating the lagged periods.
- If the goal is to capture the impact of a recent change in the environment, then indeed, it is preferable to restrict to models limited to short lags. If the model aims at capturing the evolution of the ECL and the changes in estimated PDs, including from past events, it is reasonable to include lagged events for more than a year in the PiT PD analysis.
I hope this helps...Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.