Machine Learning and AI advantage

1. The Need

Predicprobability is a key factor for the capital adequacy ting the Lapse requirement

2. The Goal

Estimating this probability as accurately as possible

3.How Do We Do It Today ?

Estimating this probability as accurately as possible

4. Our Platform

Estimating this probability as accurately as possible

MACHINE LEARNING FOR SURVIVAL ANALYSIS

Goal:

Estimate time of event for new costumers based on past costumers

Currently common solutions:

A-parametric survival curves per behavioral group
Semi parametric modeling such as Cox model.

Our suggestion:

Use ML algorithms to identify groups and construct a survival curve to each group automatically

Naïve solution:

Use available covariates to construct decision trees who identify sub groups with different survival curves and estimate them.

Can we do better?

  • Decision trees use a greedy algorithm and may end up in a local optimum.
  • Ensemble methods which combine multiple decision trees are known to have more. power and achieve higher accuracy.

Our solution:

Use Gradient Boosting Machines (GBM), a well known ensemble algorithm, to perform the task.

GBM FOR SURVIVAL ANALYSIS

Objective : Minimize the loss function

  • Define split criteria following the same logic
  • Tune the parameters using Grid Search

SIMULATIONS

Perform extensive simulations to evaluate suggested solution:

  • Over 15 simulated data sets, 10 potential covariates (varying subset of which are influential), varying time distribution and censoring.
  • Comparing to Cox PH model in terms of survival curves and parameter estimations.

Results in a nutshell:

  • GBM constantly outperform Cox in parameter estimations.
  • GBM’s survival curves are constantly closer than Cox’s to truth data.

GBM output of covariance’s importance metric supply valuable insights to the modeling process and the influential subset of covariates

ASSUMPTION 01 RELAXED

Non-Linear Dependency On The Covariates

Assume non-linear dependency on the covariates. How ?

The more data we have the better our approximation can get
Very flexible and complex models requires a lot of data and a lot of computing power

ASSUMPTION 02 RELAXED

Proportional Risk

Yes, but it is more problematic for various reasons:

Good News

In the last few years some people have done this and reported encouraging results

Scroll to top