Scoring & Segmentation - A Machine Learning Approach to Classification
The challenge presented by the client in the third-party collection industry was to find a way to better utilize their finite resources and accelerate cash flow in their collection practice. Massena proposed implementing a machine learning-based payer propensity model to address this challenge. This model would utilize historical data from the client's collection process, such as account information, payment history, and demographic data, to predict the likelihood of a delinquent consumer making a payment before any resources were expended in the collection process.
The implementation of this model involved several steps, such as data preparation, feature engineering, model selection and training, and evaluation:
The data was cleaned and preprocessed to remove any missing values or outliers.
Data points that could affect the payment likelihood were identified and engineered. The appropriate machine learning algorithm was then selected and trained using this preprocessed data.
The model was evaluated using various metrics such as accuracy, precision, recall, and F1-score to ensure it met the client's requirements.
The final model developed was a binomial classifier with a true positive rate of 88%, which means it was able to successfully predict 88 out of every 100 real payers before any efforts or resources were expended. This model was robust, easily deployable, and could predict the likelihood of immediate payment for each delinquent account. Additionally, the model produced an overall probability score that changed dynamically in response to ongoing interactions between the organization and the consumer.
The implementation of this model had a significant impact on our client's operations, leading to cost savings of $1.7 million per year and revenue growth of $4.7 million. This represented a major shift in how our client executed their operations. They could now prioritize their collection efforts on the accounts with the highest payment likelihood, resulting in increased efficiency and profitability. The model also provided valuable insights into the factors that affect payment likelihood, which the client could use to further optimize their collection process in the future.