In this blog post, our chief scientist Alexander Efremov PhD. is discussing the application of direct and recurrent artificial neural networks (ANN) in some methodologies for credit risk models.
Inputs of ANN could be the available individuals’ characteristics, like age, income, marital status, credit bureau data, etc. The outputs are the probability the applicants to have a good performance as loan-holders, the individuals’ response to some actions, etc.
Artificial neural networks /ANN/ have many applications. A new area for ANN is the credit risk and market strategies optimization. The main reason for not paying enough attention to these mathematical representations in the mentioned fields is that ANN-s cannot be verified with the available apriori information about the investigated system. Such verification is needed, as the uncertainty level in the real-life data is very high (in marketing and finance), which may cause illogical relations between the factors and the outputs. An example of model verification with apriori information is the following. It is preliminary known that the factor ‘annual income’ is positively correlated with the output ‘probability of good’, and ‘county court judgment’ is negatively correlated with that output. When we work with linear or logistic models (traditionally used for credit scoring) it is easy to check, if the model preserves these expected trends. What is only needed is to check the sign of parameters’ estimates (it should be positive for ‘annual income’ and negative for ‘county court judgment’). But when work with ANN, such a conclusion, based on the network parameters cannot be made. The reason is that generally, ANN parameters (weights) have no physical meaning.
In some countries (like USA), the rejected credit applicants may require from the financial institution the 5 major reasons they to be rejected. This information comes straightforward from the linear/logistic model estimates but again, if ANN is used, such information cannot be extracted.
Nevertheless, ANN is a powerful tool, which can be used at different stages, during the model development. One of the main advantages of the networks is that they can be used for a representation of strongly non-linear systems. This is the case in the credit risk and in the market strategies optimization, where the investigated system is the individual. ANN can be used for data cleaning, building of propensity models, demand models, “fast and dirty” models, assisting the final model development, etc.
Applications for credit risk assessment
A frequently used technique in the system identification is to generate “fast and dirty” models at the stage of data preprocessing and more precisely for data cleaning. The idea is to use a model, which estimates the possible values of missing data or to fill statistically correct values after shaving outliers. At this stage of the overall system identification cycle, the analyst does not want to spend much time in building such models (this is the reason the models to be named “fast and dirty”). Usually, there are no strict requirements these system representations to be logical and to fit with the available apriori information.
Also, it is well known that individuals’ behavior is strongly non-linear w.r.t. some factors. For these reasons, ANN-s are very suitable for data cleaning. If a linear or logistic model is used and the variables are continuous, some non-linear input-output transformations have to be applied and after that, an analysis should be performed in order to select the most appropriate transformations, which improve the model accuracy. But ANN-s don’t require such actions, which saves time. On the other hand, keeping in mind that the networks are capable to manage with unknown non-linearities, it is expected that the resulting fast and dirty ANNs will provide more accurate predictions.
Other application of ANN is when segment the total population. There are subpopulation analyses based on fast and dirty models. Here the goal is to answer the question: “How different are some segments?” before to start building models on the corresponding subpopulations. The main two types of scorecards in the credit risk assessment are related to the application for a credit (application scorecards) and how people service their loans (behavioral scorecards). First models explain only the static part of the individuals’ behavior, as the data doesn’t contain information about the dynamics. They are used to predict the risk level, associated with credit applicants. On the other hand, the behavioral models represent the dynamic performance of the accepted applicants. Hence for application scorecards, feedforward (fast and dirty) ANN-s can be used, but for behavioral scorecards, recurrent ANN-s would be preferable.
ANN-s have important advantages, compared with the linear and logistic models, used in marketing and finance industries. The major advantage of the networks is their ability to represent non-linear behavior and as it is well known, the individuals’ behavior is strongly non-linear.
ANN-s can be used for modelling of many, completely different aspects of the investigated system, but the training process remains the same. For instance it is not necessary at the preprocessing stage to apply non-linear transformations on the input-output processes. This is very important, when the system identification cycle is applied without (or with limited) human intervention.
Now-a-days, a huge amount of data, regarding the individuals’ performance is available. From this perspective, the above mentioned potential of ANN-s, combining with the possibility of an automated model development is a premise for their successful usage in solving of big data problems.
