Page tree

Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

But with regard to ...

  • Staging,
    IFRS9 IFRS 9 differentiates between 3 stages depending on the significance of the deterioration in credit quality. Depending on the stage assigned, IFRS9 IFRS 9 calls for the consideration of 12-month or lifetime ECLs.
    As far as early warning is concerned, the differentiation between the 3 stages does not have a sufficient degree of fine granularity. For example, it makes a difference in the probability of default from the perspective of the individual deal if a retail customer is delayed with a due payment for 5 days, 2 weeks or even 4 weeks. Even the probability of success of forbearance or modification measures depends on the right timing which is not necessarily provided by regulatory approaches.

  • Segmentation,
    Conventional approaches work with a rule-based segmentation, applying manually captured settings to identify similar credit risk characteristics using an excerpt of the customer DNA and some product and deal-specific parameters. The calculation of lifetime expected credit losses using parameters based on portfolio/segment, defined using high level similar credit risk characteristics, is fine.  However, e.g. for significant deals, the use of regulatory segments to decide on a credit risk adjusted price is questionable. Due to the rough granularity of segments, an entity might not make a deal because the interest offered was too high or it might conclude a deal where the interest does not cover the credit risk. 

...

Conventional approaches are sufficient and accepted for regulatory requirements. Of course, they can be generally used for business analysis too.
However, in order to improve business decisions, an approach should be applied that considers the full individual customer DNA and the entirety of the deal parameters. Machine learning considers all this during the training of the neuronal network. Any inaccurancy caused by manual staging and manual segmentation can be avoided.


Image RemovedImage Added                                
         

...