2.7.1 | The objective of this section is to draw attention to the key challenges and minimum expected practices to ensure that institutions develop effective rating models. The development of retail scorecards is a standardised process that all institutions are expected to understand and implement appropriately on large amounts of data. Wholesale rating models tend to be more challenging due to smaller population sizes and the complexity of the factors driving defaults. Consequently, this section related to model construction focuses on wholesale rating models. |
2.7.2 | Wholesale rating models should incorporate, at a minimum, financial information and qualitative inputs. The development process should include a univariate analysis and a multivariate analysis, both fully documented. All models should be constructed based on a development sample and tested on a separate validation sample. If this is not possible in the case of data scarcity, the approach should be justified and approved by the validator. |
2.7.3 | Quantitative factors: These are characteristics of the obligors that can be assessed quantitatively, most of which are financial variables. For wholesale rating models, institutions should ensure that the creation of financial ratios and subsequent variable transformations are rigorously performed and clearly documented. The financial variables should be designed to capture the risk profile of obligors and their associated financing. For instance, the following categories of financial ratios are commonly used to assess the risk of corporate lending: operating performance, operating efficiency, liquidity, capital structure, and debt service. |
2.7.4 | Qualitative subjective factors: These are characteristics of the obligor that are not easily assessed quantitatively, for instance the experience of management or the dependency of the obligors on its suppliers. The following categories of subjective factors are commonly used to assess the risk of corporate lending: industry performance, business characteristics and performance, management character and experience, and quality of financial reporting and reliability of auditors. The assessment of these factors is generally achieved via bucketing that relies on experts’ judgement. When using such qualitative factors, the following principles should apply: |
| (i) | Institutions should ensure that this assessment is based upon a rigorous governance process. The collection of opinions and views from experienced credit officers should be treated as a formal data collection process. The data should be subject to quality control. Erroneous data points should also be removed. |
| (ii) | If the qualitative subjective factors are employed to adjust the outcome of the quantitative factors, institutions should control and limit this adjustment. Institutions should demonstrate that the weights given to the expert-judgement section of the model is appropriate. Institutions should not perform undue rating overrides with expert judgement. |
2.7.5 | Univariate analysis: In the context of rating model development, this step involves assessing the discriminatory power of each quantitative factor independently and assessing the degree of correlation between these quantitative factors. |
| (i) | The assessment of the discriminatory power should rely on clearly defined metrics, such as the accuracy ratio (or Gini coefficient). Variables that display no relationship or counterintuitive relationships with default rates should preferably be excluded. They can be included in the model only after a rigorous documentation of the rationale supporting their inclusion. |
| (ii) | Univariate analysis should also involve an estimation of the correlations between the quantitative factors with the aim to avoid multicolinearity in the next step of the development. |
| (iii) | The factors should be ranked according to their discriminatory power. The development team should comment on whether the observed relationship is meeting economic and business expectations. |
2.7.6 | Multivariate analysis: This step involves establishing a link between observed defaults and the most powerful factors identified during the univariate analysis. |
| (i) | Common modelling techniques include, amongst others, logistic regressions and neural networks. Institutions can chose amongst several methodologies, provided that the approach is fully understood and documented internally. This is particularly relevant if third party consultants are involved. |
| (ii) | Institutions should articulate clearly the modelling technique employed and the process of model selection. When constructing and choosing the most appropriate model, institutions should pay attention to the following: |
| | (a) | The number of variables in the model should be chosen to ensure a right balance. An insufficient number of variables can lead to a sub-optimal model with a weak discriminatory power. An excessive number of variables can lead to overfitting during the development phase, which will result in weak performance subsequently. |
| | (b) | The variables should not be too correlated. Each financial ratio should preferably be different in substance. If similar ratios are included, a justification should be provided and overfitting should be avoided. |
| | (c) | In the case of bucketing of financial ratios, the defined cut-offs should be based on relevant peer comparisons supported by data analysis, not arbitrarily decided. |