Evaluating the Impact of Feature Engineering on Auto Insurance Claim Prediction Models

Predicting automobile insurance claim is essential for risk assessment, premium calculating, and fraud detection. All of which help ensure the profitability and stability of insurance companies. This study introduced a novel engineered feature, the Collision Index, designed to quantify localized automobile collision risk, which is aggregated into an insurance claim dataset to evaluate its impact on model performance. Four machine learning models, Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine and Neural Network, were trained using the dataset with and without the engineered feature. Their F1 scores are compared using the paired Student’s t-test. Results indicate that only the Random Forest model witnessed a statistically significant improvement with the inclusion of the collision index at 0.05 significance level. The other models fail to observe significant performance gain, or imply a decisive conclusion due to computational constraints. Limitations of this paper include the imbalanced dataset, the estimated nature of collision index and possible incorrect assumption made during paired t-test.