We offer highly sophisticated credit risk assessments based on our machine learning model XGBoost. Multiple studies have also demonstrated the superior performance of our machine learning model when compared to other methods. For example, our machine learning model offers solutions to multiple problems that the logistic regression faces, including:
- Importance of the variables stays constant for all companies (see example situation below)
- Increasing the number of explanatory variables can lead to more unstable predictions
- Possible outliers’ influence on the model can get substantial
We have also demonstrated the accuracy of our machine learning model compared to a traditional logistic regression model. You can see the accuracy comparison by clicking here. Additionally, you will find literature references and ROC-AUC values about different methods in various academic studies from our credit risk presentation (page 13) by clicking here.
However, a challenge with machine learning models is their lack of explainability, as they operate as black boxes, making it unclear to users what occurs within the model. For example, with logistic regression models we would always know the impact of each explanatory variable on the final risk value since the weights are constant.
We have tackled this problem by introducing a scatter plot visualization accompanied by an automatically generated text, which elaborates on the reasoning behind the decisions of our bankruptcy risk prediction. See example below.
VISUALIZATION & AUTOMATIC TEXT
Below, you will find an automatically generated text, accompanied by a visualization on the right. The text has been prepared for a company that is situated with the highest-rated companies even though our credit risk model marks it as a high-risk company:
The company has very high profitability and solvency. For example, in 2020, the ROA-% of Company X was 23.1 % and the equity ratio was at 81.7 %. The net sales in 2020 were 845 kEUR which represents a growth of 13.1% from the year before. While the company has excellent figures in these aspects, the credit risk model has rated the company much lower than other companies with similar profitability and solidity. The higher credit risk is a result of the following weaknesses identified by the model:
Increasing current loans receivable: From 2016 to 2020, current loans receivable grew from €22k to €186k, indicating that the company is lending out more money, which could result in bad debt if borrowers default.
Low cash and cash equivalents: The company has consistently low cash balances, with only €5k in cash at the end of 2020, which may make it difficult to cover short-term obligations or unexpected expenses.
High non-interest-bearing liabilities: In 2020, non-interest-bearing liabilities reached €68k, putting pressure on the company’s liquidity and potentially increasing bankruptcy risk if they are unable to pay off these liabilities.
Based on the above-mentioned factors, our credit risk model has assessed that the company has a high bankruptcy risk of 0.947 %, which corresponds to a credit rating of BAA (poor).
ADVANTAGES OF MACHINE LEARNING MODELS
The primary advantage of machine learning models lies in their ability to utilize dynamic weights for various variables. Unlike a simple regression model, which relies on a fixed polynomial equation, machine learning models comprise numerous decision trees. The selection of the appropriate decision tree branch is determined based on the specific situation. An illustrative example is provided below, accompanied by an image:
Company A has a very good solvency and profitability. Company B on the other hand has very poor solvency and it is unprofitable. When assessing their credit risk, these companies should have different weights for the explanatory variables like liquidity.
Here, Company A doesn’t need to have good liquidity since it is able to fund itself through its operations or by loaning money. On the contrary, Company B is losing money and can’t raise loans. The most important feature it has is its liquidity.
It can be clearly seen that varying weights are necessary for succesful credit risk assessment. Logistic regression has constant weights and thus it is unable to account for these firm-specific characteristics. Machine learning algorithms on the other hand can recognize that the significance of liquidity becomes larger with unprofitable companies and will adjust its credit ratings accordingly.