AICrowd Insurance Pricing Game - 1st Place Solution

https://www.aicrowd.com/challenges/insurance-pricing-game

Folder structure:

Preprocessing.py

Feature engineering

1. Binning
   Separate continuous variable into segments (Clipping is implicitly done too)
   Used for GLM to help capturing non-linear relationship
   
2. Interactions
   a) Population density
   b) Driver Gender combination
   c) Vehicle feature interactions
   vh_value * vh_weight
   present_vh_value (exponential decay by vh_age)
   and more...
   
3. Grouping
   Grouped Med1 and Med2 together in policy type
   
4. Transformation
   Log-transform, power transform of some continuous variables
   
5. History variable
   Historical Claim amount, Historical claim count, year since last claim, change in NCD

Training.py

Large Claim detection model:

• A XGBoost and Logistic regression model to predict whether a claim would be >3k.
  

Claim estimation model:

• I stacked 7 base models using a Tweedie GLM as the meta-learner under 5 fold CV.
  
  • Tweedie GLM
    
  • Light GBM
    
  • DeepForest
    
  • XGBoost
    
  • CatBoost
    
  • Neural Network with Tweedie deviance as loss function
    
  • Neural network with log-normal distribution likelihood as loss function (learning the mu and sigma of the loss)
    

Model.py

• The script that is used to produce prediction inside the AICrowd environment
• Pricing strategy is incorporated in the predict_premium function

The final presentation is also uploaded to this repository.