Cost-sensitive Global Model Trees applied to loan charge-off forecasting

a r t i c l e i n f o Keywords: Cost-sensitive regression Model trees Evolutionary algorithms Asymmetric costs Loan charge-off forecasting Regression learning methods in real world applications often require cost minimization instead of the reduction of various metrics of prediction errors. Currently in the literature, there is a lack of white box solutions that can deal with forecasting problems where under-prediction and over-prediction errors have different consequences. To fill this gap, we introduced the Cost-sensitive Global Model Tree (CGMT), which applies a fitness function that minimizes an average misprediction cost. Proposed specialized genetic operators improve searching for optimal tree structure and cost-sensitive linear regression models in the leaves. Experimental validation is performed on loan charge-off data. It is known to be a difficult forecasting problem for banks due to the asymmetric cost structure. Obtained results show that specialized evolutionary algorithm applied to model tree induction finds significantly more accurate predictions than tested competitors. Decisions generated by the CGMT are simple, easy to interpret, and can be applied directly. A lot of real world problems are cost-sensitive, which means that different types of prediction errors are not equally costly [5]. As a result, typical minimization of prediction errors is not the best scenario. A cost-sensitive term encompasses all types of learning where cost is considered [50,19], and different types of costs (e.g., costs of attributes, cost of instances, and costs of errors) can be distinguished. The current research focuses on a single cost for decision making, however, multiple costs are also investigated [35]. For example, in medical diagnoses, there are several types of costs that can be minimized, such as the cost of misclassification (e.g., overlooking an ill patient can be fatal in contrast to a false positive test) or the cost of treatment (e.g., financial or risk). When speculating on stock exchange, investors directly compare future gains and losses and usually give more weight to losses. Researchers show that potential gains need to be approximately twice as large to offset potential losses [51]. As a consequence, investors tend to realize their gains more often than their losses as they sell winning stocks more readily. There are many other examples for such asymmetry, such as in bankruptcy prediction [57], behavioral finances [45], expected stock returns [2], criminal justice settings [6], physician prognostic behavior [1], product recommendations [32], and so on. Cost-sensitive regression is still not adequately addressed in …