A fresh take on risk and valuation

How Finalyse can help

MACHINE LEARNING MODEL VALIDATION FRAMEWORK

How do you make sure your models are performing as expected?

Risk Process Improvement

Streamlining risk processes in order to achieve cost savings, improve output quality and ameliorate time management

Back
Article

Behavioral Scorecard with Machine Learning Components

Predicting the future is not magic, it is Artificial Intelligence - Dave Waters.

Written by Ega Skura, Senior Consultant

 

 

Introduction

Today’s bank lending process is characterised by uncertainty and volatility. At Finalyse, we help banks quantify their risks. We see credit scoring as a vital risk management tool, revealing untapped opportunities to create a healthy loan portfolio.

However, the complex regulatory landscape, the changes in business strategy, new technologies, and the uncertainty of the overall economy lead to the need of building robust scoring models. Quality scoring models help controlling credit risk, identifying opportunities, and lead to a new way of doing business. 

credit score is a numeric value that represents the creditworthiness of the customers. Credit scoring models are categorised into two different types: application scoring and behavioural scoring. Application scoring attempts to predict a customer’s default risk at the time an application for credit is made, based on information such as applicant demographics and credit bureau records. Behavioural scoring assesses the risk of existing customers based on their recent accounting transactions, recent financial information, including repayment performance, delinquencies, credit bureau data, and their overall relationship with the bank. By identifying the riskiest clients, the bank can take preventive actions to hedge itself against future potential loss.

Machine Learning is a method of teaching computers to parse data, learn from it, and then make a decision or prediction regarding new data. Machine Learning overlaps with statistical learning as both attempt to find and learn from patterns and trends within large datasets to make predictions.


This article discusses the benefits of applying advanced analytics in the development of behavioural scoring models. It investigates how Machine Learning techniques can be used to model the behavioural scores of consumers in each step of the development and model assessment. Some concerns regarding the usage of Machine Learning in behavioural scoring model are addressed. 
 

Motivation of using Machine Learning in modelling

Machine Learning offers various algorithms and methods that can be used during model development. There is a wide variety of Machine Learning applications related to feature selection, model fitting techniques, and model quality assessment during model development. The advantages of applying machine learning approaches consist mainly of:

  • Increasing the efficiency and the effectiveness of model development and assessment
  • Capturing the complex non-linear relationships in the respective dataset
  • Facilitating the decision-making process based on the clients' risk categorisation

Enhance the efficiency and effectiveness

Capture non-linear relationships in data

Faster decision making

 

Figure nr.1: Main benefits of Machine Learning application in modelling.


 

The following chart summarises the contributions of machine learning in the development of the behavioural scoring model.

 

Figure Nr.2: Benefits and use of the Machine Learning techniques in behavioural scoring.


 

Behavioural Scoring development and machine learning usage

 

The development of the Behavioural scoring consists of the following steps: Data collection and sampling; Exploration and treatment of the data; Classification/Grouping; Model selection and model testing/evaluation.

 

 

Figure Nr.3: Behavioural scoring development steps.

 

Each step relates closely to those that precede and follow it and has an impact on the overall model developed, so it is important to analyse each aspect involved. Based on the implication of Machine Learning methods in Behavioural scoring model the following part explains the interaction based on 4 main domains of model development: sample creation, data engineering, model engineering, and model evaluation. 

 

 

How to build a Machine Learning Behavioural Scoring Model?

 

Domains

ML components

Sample Creation

 

Big data gives the opportunity to build a good model. When more data is available, increasingly complex data structures and relations can be found. ML techniques are powerful in handling complex and big data.

Feature engineering

Usage of the ML approaches (Wrapper, Filter, and Embedded) for feature selection. Usage of ML methods for feature binning (K-means, hyperparameter optimisation).

Model development

Usage of powerful Machine Learning algorithms produces models with high accuracy. Using explainability techniques (LIME, SHAP), ML models are not a ‘black’ box.

Model Assessment

Statistical metrics (Confusion Matrix, Gini Coefficient, Kolmogorov Smirnov Chart, Root Mean Squared Error, etc.) are measured based on the type of the ML model. Accurate model evaluation should be based on cross-validation.

Table Nr.1: Behavioural scoring model’s domains and Machine Learning components.

 

Sample creation

  • Data should be representative of the credit portfolio. There are two approaches to behavioural scoring: techniques that use static characteristics (features assessment provided after 12 months) about the customer’s past performance and techniques, which incorporate dynamic aspects (features assessment provided for months 3, 6, 12, or 24). Features indicating the customer delinquencies and the repayment history should be calculated. 
  • Sampling is related to the longitudinal aspect of the data used in behavioural scoring. A sample of customers is selected so that their repayment behaviour on either side of an arbitrarily chosen observation point is available. The period before the observation point is often termed as the performance window. Data on the customers’ performance during this time is structured into features that are used by the behavioural scoring system to distinguish between customers likely to repay their loan and those likely to default on their financial obligations. In this stage, the target variable is defined. Machine Learning can be helpful with the advanced analysis for finding the optimal period length of the target feature, the longitude of the observation, and the outcome, in order to eliminate the seasonal behaviour of the portfolio and have a reliable model. This analysis consists of producing multiple ML models for each target feature, observation and outcome length, assessing the performance of each model, and deciding the period length that best fits with the data and the product lifecycle.

 

Benefits of using Machine Learning in Sample creation:


Sample creation consists of collecting data, calculating features, defining the performance window length, the outcome length, and the target feature. Machine Learning helps in making the optimal choices based on given data.


 

Feature engineering

  • Exploration, treatment, and quality assurance of the data - this stage consists of the analysis of the features (descriptive statistics), missing values, and outliers. The treatment of missing values and outliers is extraordinarily important; it leads to the choice of the optimal Machine Learning algorithm. Whichever treatment transforms data to make learning easier leads to a significant improvement in the final performance.
  • Features selection: it is also important to reduce the number of input variables exclusively to those that are believed to be the most useful for predicting the target variable. Feature selection primarily focuses on removing non-informative or redundant predictors from the model. There are several Machine Learning approaches, which can be used for this purpose. 

 

Wrapper approach

  • Tries to search for well-performing subsets of features using a machine learning algorithm.
  • Common ML algorithms used in feature selection can be recursive feature elimination (starts with all features and removes features until the desired number remains), forward feature selection, backward feature elimination.
 

Filter approach

  • Select features in any machine learning algorithm by selecting subsets of features based on their relationship with the target.
  • Some of the filtering methods are: statistical methods, feature importance methods, intrinsic, decision trees (project input data into a lower-dimensional feature space).

 

Embedded approach

  • Uses algorithms that have built-in feature selection methods.
  • Regularization methods are the most used embedded methods, they penalize a feature by assigning coefficient threshold. embedded algorithms include: LASSO, elastic net, ridge regression.

 

 

 

 

The principal component analysis is a ML technique for feature extraction —it creates new independent features, where each new independent feature is a combination of the “old” independent features. It determines the lines of variance in the dataset so that the first principal component has the maximum variance, thus helping in overcoming the overfitting issue by reducing the number of features. Although principal components try to cover the maximum variance among the features in a dataset, if the number of principal components is not selected with care, we may miss some information as compared to the original list of features. The proper Machine Learning approach is chosen based on the data availability.

  • Binning/Grouping - relates to the grouping of features into bins/clusters. Grouping/binning the features in several categories is essential for building a robust model with fewer coefficients. There are several methods of statistical grouping for numeric features: Equal interval grouping, Equal frequency grouping, Weighted of Evidence and Information Value. For grouping the categorical features, the Chi-Square method can be used.  

 

Benefits of using Machine Learning in Feature Engineering:


Feature engineering consists of creating, selecting, and grouping features that contribute to a representative dataset creation by increasing the predictive power of machine learning algorithms.

­

Model building

Train the model - this step consists of building several behavioural models and selecting the best model that fits the data and the business strategy. The learning algorithm finds patterns in the training data such that the input parameters correspond to the target. The output of the training process is a Machine Learning model that you can then use to make predictions. Machine Learning offers various algorithms, which produce models with high accuracy in the prediction. Different Machine Learning approaches could be used during the modelling part. The graph below presented the accuracy of different Machine Learning models based on a Polish corporate bankruptcy dataset. 1

Source: Dissertation ‘Analysis of financial credit risk using Machine Learning’, Aston University, Birmingham, United Kingdom
Graph Nr.1: Accuracy rate in different Machine Learning model

 

Benefits of using Machine Learning in Model Engineering:


Machine learning offers powerful algorithms leading to robust models.


 

Model Assessment

Evaluating the model: in this final stage, we analyse qualitative and the quantitative aspects of the model.

  • Quantitative analysis statistically proves that the performance of the model remains the same in the test dataset. The model selected should be tested in terms of consistency and accuracy. ML-based methods can provide different metrics for measuring the model performance, some of them are shown in the table below:

 

Metric

Regression models

Classification models

Unsupervised models

Other

Mean Squared Predictor Error

 

 

 

 

Mean Squared Absolute Error

V

 

 

 

Adjusted R Square

V

 

 

 

Precision-Recall

 

V

 

 

Receiver Operating Characteristic-Area Under Curve

V

V

 

 

Accuracy

V

V

 

 

Log-Loss

 

V

 

 

Rand Index

 

 

V

 

Mutual Information

 

 

V

 

CV Error

 

 

 

V

Heuristic methods to find K

 

 

 

V

BLEU Score (NPL)

 

 

 

V

Table Nr.2: Performance metrics based on the Machine Learning model.

 

Cross-validation is a statistical technique used during the performance assessment of a Machine Learning model. Cross validation gives a comprehensive measure of the model’s performance throughout the whole dataset by dividing it into n subsets. An average is calculated over error estimation s over all n trials to get the total effectiveness of our model. Every data point needs to be in the validation set exactly once and needs to be in a training set n-1 times.

  • Qualitative analysis is related to the explainability of the model. The two most used techniques are LIME (model-agnostic technique which attempts to understand the model by perturbing the input of data samples and looks at how the predictions change) and Shapley (a united approach for explaining the output of any Machine Learning model by assessing the marginal contribution of a feature value). The following picture schematically presents the Sharpley functionality by finding the contribution of each feature (input) in the model in order to understand which feature contributes more and which feature contributes less to the total customer scoring. Find out more about the Machine Learning model explainability.

 

 

 

Benefits of using Machine Learning in Model Assessment:


Model Evaluating helps in ensuring that the model performs in another dataset. To avoid overfitting cross-validation methods should be used. Also, Machine Learning techniques can be used for qualitative analysis.

 


 

Prudence in using Machine Learning in Behavioural scoring

The selection and application of the proper Machine Learning algorithm which best fits with the data, the nature of the portfolio, the expected outcome, and the target feature are crucial. The usage of Machine Learning algorithms sometimes leads to the creation of overfitting models. There are two main types of error in Machine Learning models:

  • Bias error - caused by assumptions. It appears when there is a substantial difference between the expected average value estimated and the actual value. The expected estimated value is the average of values acquired when the model is trained multiple times.
  • Variance error - caused by fluctuations in data, for example, noise. High variance can cause a model to be prone to overfitting. And training the same model on different data sets will yield different results when the variance is high.

 

In Machine Learning, every model has some bias and some variance. Decreasing variance increases bias. And decreasing bias increases the variance. Some models are more prone to different types of error as summarised in the following scheme.

 

Linear algorithms are prone to high bias making them easy to learn and understand but generally less flexible. They have lower predictive performance on complex problems that fail to meet the simplifying assumptions of the algorithm.

Changing model parameters can help mitigate one type of error but it cannot solve both. Some types of hybrid  models may have however the ability to mitigate both types of error. As the hybrid2 model for building behavioural scoring can be used to find an appropriate method for the feature and classifier selection, in order to use an optimal classifier and feature subset.


[2] Zhang, W., H. He, and S. Zhang, A novel multi-stage hybrid model with enhanced multi-population niche genetic algorithm: An application in credit scoring. Expert Systems with Applications, 2019. 121: p. 221-232


It is important to use cross-validation methods to eliminate the overfitting issue. Using cross-validation, the dataset can be spit n times. The next step is to train each dataset, calculate the accuracy for each dataset, and an average of the metrics. This method requires somewhat more computations than regular splitting. But it returns much stronger evaluation metrics. Before selecting any Machine Learning method, it is worth knowing the pros and cons of applying them in Behavioural scoring modelling. The following table summarizes merits of some Machine Learning algorithms:

Machine Learning Algorithm

Description

Pros

Cons

Logistic Regression

Estimation of the relationship between variables to predict the future outcome.

Outputs have probabilistic interpretation, and the algorithm can be regularized to avoid overfitting. Logistic models can be updated easily with new data using stochastic gradient descent.

Logistic regression tends to underperform when there are multiple or non-linear decision boundaries. They are not flexible enough to naturally capture more complex relationships.

Decision Trees

Decision trees are rule-based learning methods.

 

Easy to implement and understand as they require data normalization and the model is not affected by missing values.

 

A small change in the data can cause a large change in the structure of the decision tree causing instability.

 

Support Vector Machine

SVMs finds optimal hyperplanes that best separate the dataset. A hyperplane is chosen such that the ‘margin’ between the plane and data points of different classes on either side is maximized.

High accuracy model with a high number of features.

The algorithm is prone to over-fitting, if the number of features is significantly greater than the number of samples. Also, SVM models require a big dataset.

Neutral Network

NNs are sets of algorithms which try to recognize the patterns in the data and help cluster and classify the data.

Neural networks are good to model with nonlinear data with a large number of inputs. It works by splitting the problem of classification into a layered network of simpler elements.

Neural networks are black boxes, too often, we cannot know how much each independent variable is influencing the dependent variables.

Gradient Boosting

GB builds an ensemble of successive trees with each tree learning and improving on the previous.

No data pre-processing required - often works great with categorical and numerical values as is.

Many parameters influence the Behavioural of the approach (number of iterations, tree depth, regularization parameters, etc.). This requires a large grid search during features allocation and transformation.

Table Nr.4: Pros and Cons in using Machine Learning algorithms in modelling. 

 

Conclusion

Machine Learning techniques can bring great added value to building statistical models. This document aimed to show some concepts of how Machine Learning components can be used in the context of behavioral scoring model and to shortly describe some techniques that can be employed. Each technique should be applied to conform to the data structure, size, and composition. The implication of the Machine Learning methods can be used in each step of the behavioural model development to ensure an adequate process for the overall model development. This article shows that the usage of the advanced Machine Learning methods could result in behavioural scoring models with high accuracy in predictive power but every aspect of the model development should be considered carefully in order to adapt the proper Machine Learning algorithm. Knowing the pros and cons of each algorithm can help in choosing the best approach in line with the model purpose and model usage.