Forecasted purchase of caravan insurance using Random Forest Classifier
Our goal was to build a machine learning model that can accurately and reliably predict whether someone in a given postal code will purchase caravan insurance. Such a model could be used to market to members most likely to purchase insurance.
The table below summarizes the machine learning workflow. After cleaning data and adding new features, such as number of policies purchased, correlated variables and low-variance variables were removed. Then, data was truncated using Singular Value Decomposition (SVD) for computational efficiency and improved performance. Lastly, cross validation with 5 folds was applied to the following models : Logistic Regression, Random Forest Classifier, ADA Boost, and XG Boost. Random Forest showed the best results with a weighted f1-score of 0.11.
Achieved F1-Score of 0.19 on test data, with a modest recall at 0.54 but a low preciion of 0.11. This model is not yet ready for real-world application, but taught us how to optimally transform data as well as train and compare machine learning models.
