Auto-Insurance-Claim-Predictor

Python ML project using scikitlearn,pandas,numpy,matplotlib

Problem Statement:

Build a Machine learning model to predict the probability that owner will initiate an auto insurance claim in the next year.

Most companies charge a flat premium to the customers irrespective of their risk for filing an insurance claim. Inaccuracies in the insurance company’s claim predictions raise the cost of insurance for good drivers and reduce the price for bad ones. Our project will help the insurance company in following ways:

Affluent customer: Company can attract the good drivers if it is doing the fair pricing. Loss ratio: Company can avoid specific customer/policies if they are at high risk of filing claim which in turn decrease loss ratio. Fair pricing: Company can charge the premium to the customers by their risk, and accurate prediction will allow them to tailor their prices further. Claim forecast: Claim is proportional to the number of risky customers, so company forecast the number of claims it could get next year which will help them to manage their fund better

As per industry estimate 1% reduction in the claim can boost profit by 10%. So, through the ML model, we can identify and deny the insurance to the driver who will make a claim. Thus, ensuring reduced claim outgo and increased profit.

Analysis Approach

Data Pre-Processing

• Load the Data into DataFrame
• Inference about the data
• Check the distribution of target variables
• Split features into Categorical,Binary,Ordinal,Interval
• Found Outleirs in continuous variables
• Feature Engineering
  • Handling Missing Values
  • Found Outliers after filling missing values
  • Dropped Id column
• Data Visualization
  • Inferences about continuous,ordinal,binary variables through Visualization
  • Univariate
    • Distribution of continuous variables - Histogram, Distplot, boxplot
  • Bivariate
    • Categorical variables vs count for ordinal features - barplot
    • Categorical variables vs count vs target distribution - barplot
  • Multivariate
    • Correlation Analysis - Level of Correlation for interval,ordinal features
• Keep only 1 column for columns having high correlation
• Balanced the data by OverSampling(Smote algorithm)
• One-hot encoding/dummification of the categorical variables
• Saved CSV Files for data with and without Encoding

Modeling

• Split the data into training and testing using train_test_split
• Standardize the data by applying fit on training data and transform on train and test data
• Classification algorithms shown below are used to build model
  • Logistic Regression
  • Support Vector Machines(Linear,Poly)
  • KNN
  • Gausiian Naive Bayes
  • Decision Trees
  • Random Forest,
  • XGBoost,AdaBoost,GradientBoost are used to build the model
  • MLPClassifier
• Model Evaluation Metrics
  • accuracy
  • precision_score
  • recall_score
  • f1_score
  • confusion matrix
  • classification report are displayed
• Model Selection
  • Same results are observed for tree based models with and without OneHotEncoding categorical features while results improved for rest on OneHotEncoding
  • Feature Importance from the model is also displayed
  • Ensembling models like RandomForest and Boosting gave the best results among all the algorithms used