Motor Price Elasticity & Catastrophe Modelling (Frequency & Severity) Overview

Very much in its infancy

Hello!

This is a work in progress - all the code can be found in my

with guides (in development) on how to replicate the projects

The Projects

The Projects!

1. Conversion modelling with a randomly generated data (.csv) of motor data

to predict sales, conversion probability & elasticity

2. Catastrophe - specifically to model frequency & severity

A blend of real world data from IRIS & some Randomly generated fields

Random data - Price of commercial property, premium etc.

3. Both data sets are set to grow as the project progresses

And as the models are trained (plus some feedback)

---

Motor Conversion Model

1. Goal - with our synthetic dataset can we find out the conversion rate based on a range of different variables

2. What is the price elasticity based on each customer based on the given factors - age, location, vehicle type, number of convictions etc, etc, etc

3. Modelling - Train the data, split it into a Train, Validation & Test set. We will be training the model using a blend of R & Python

   1. R to generate the random data

   2. Python to do the modelling

   3. Model will be constructed using LightGBM

   4. Parameter tuning to be done using Bayesian Optimization instead of brute force grid search

   5. Classification model

   6. Metrics & Model performance evaluation

      1. Log Loss

      2. Actual vs Expected (AvE)

      3. SHAP

      4. Potentially PDPs

4. Visualisation - for the moment Plotly & SHAP

   1. Potentially a Shiny App

   2. Dash

   3. Other Open Source software's in research

---

Tech Stack

R to generate the data

• Pull the IRIS data set for quake data (Catastrophe Freq & Sev modelling)

• Generate randomised datasets (Commercial lines of business - Conversion)

Python to model the datasets produced (above)

• Pandas Profiling - visualise the distribution of each factor

• For modelling the main modules we depend on are

  • LightGBM (Training the model)

  • Bayesian Optimization (Hyperparameter tuning)

  • SHAPley values (determine the effect each single factor has on the model)

  • PDP

Plotly (R or Python)

• Visualise our data anything from

  • Average VS Expected plots

  • Performance by factor

  • Price Elasticity

  • Etc

Leaflet(R or Python)

• Mapping