The idea that a model is only as good as the data it is fed is a truth observed in credit risk modelling; indeed data sat at the top of Bank of England’s concern for IFRS 9 models.
Calibrating machine learning models is however not trivial and in particular for wholesale banks where the customized nature of lending causes lack of data it is even more important to focus on the machine learning model calibration, to avoid for example overfitting.
Watch the recording below.
1. Preparation of data
Three general areas will be discussed in the webinar:
- Transformations:For example non-linear relationships into linear or weight-of-evidence.
- Outliers/Missing values:Need to decide how to deal with deviating data points, considering amongst others data clustering as well as the strategy for missing values.
- Segmentation/bucketing of data:Reducing granularity which is important to avoid overfitting.
2. Selection of suitable machine learning models for the modelling problem
3. Calibration of models
- How to avoid overfitting (cross-validation, regularizations, more/less data, ensambling).
4. Evaluation of results:
- Which performance measures to use to evaluate certain types of machine learning models.
Thomas Aldheimer, Data Scientist, FCG
Laurent van Malderen, Retail Credit Risk Management, AXA Bank
Louis Brown, Head of Credit Risk Modelling, Investec Bank