Vehicle Insurance Cross-Sell Model
Uploaded by
Shubham Wadhwa 23Vehicle Insurance Cross-Sell Model
Uploaded by
Shubham Wadhwa 23Common questions
Powered by AIThe classification trees faced limitations as most response points were 'No', leading to skewed results where all nodes led to the same conclusion (0 or No response). This model's skewness towards the majority class prevented it from effectively distinguishing between respondents, a contrast to the logistic regression model which provided more nuanced results despite having a lesser error rate .
The factors identified as having significant differences between respondents and non-respondents included Gender, Age, Driving Licence, Previous Insurance, Vehicle Age, Vehicle Damage, and Annual Premium . These variables demonstrated noteworthy distinctions in their mean values when comparing those interested and not interested in vehicle insurance .
The stepwise method helped pinpoint the most significant variables by iteratively testing and adding or removing predictors based on their statistical significance and contribution to model accuracy . This process streamlined the model to only include relevant predictors such as Gender, Age, Driving Licence, Previously Insured, Vehicle Age, Vehicle Damage, and Annual Premium, refining the analysis and improving the prediction efficiency .
The team favored the logistic regression model because it produced a lower error rate in predictions compared to the classification tree. Although a classification tree was initially considered, its skewed outcomes with most leaves resulting in 'No' responses made logistic regression, with a total error rate of 36%, more reliable for their specific prediction needs .
Driving Licence was found to have a significant impact on the prediction of interest in vehicle insurance. The data analysis showed a noticeable difference in the mean response for insured as compared to non-insured individuals, highlighting that having a driving licence is a pivotal factor in insurance interest predictions .
The rationale was twofold: Firstly, there were approximately 150 different codes, making their conversion into factors add complexity without necessarily enhancing predictive power. Secondly, the presence of codes with minimal representation posed a challenge in determining a meaningful threshold for combining them. Thus, these reasons contributed to the decision to exclude them from conversion for clarity and model simplicity .
The data analysis team utilized the dataset effectively without needing to address missing values since the dataset did not have any missing entries to begin with . This allowed them to proceed directly with model building processes such as developing logistic regression and classification tree models without having to implement any data imputation strategies .
The team initially built an all-inclusive model and used a stepwise method to identify the best set of variables. They ultimately included Gender, Age, Driving Licence, Previously Insured, Vehicle Age, Vehicle Damage, and Annual Premium, deciding to exclude variables like the policy sales channel and region code due to their complexity in terms of the large number of unique values and their limited added value in improving predictions . The challenge with policy sales channel and region code was that they were represented numerically, with a high number of different codes which made them cumbersome and not beneficial for model clarity and accuracy .
The Hoslem test was not considered reliable due to the size of the dataset, which exceeded 25,000 records. It's known that for very large datasets, the Hoslem test tends to indicate poor fit; hence it was deemed not a suitable criterion for assessing the goodness-of-fit of the logistic regression model in this analysis .
The exploration of ROC curves served to determine the optimal threshold for classifying the response in logistic regression. By plotting the ROC curve, the team could analyze the trade-offs between sensitivity and specificity at various thresholds, identifying the best cutoff point to maximize model accuracy . The analysis concluded with the identification of a specific threshold that balanced sensitivity and specificity, crucial for improving predictive performance on test datasets .