The SOA likes to break links, so I'm keeping track of sources here

2010 experience study: https://www.soa.org/research-reports/2010/research-shock-lapse-report/

The dataset provided has age as a factor, with bands as levels. The 2014 report https://www.soa.org/experience-studies/2014/research-2014-post-level-shock/ has average values for each band, so we can use those.

Describe and justify the weighted RMSE measure and how the train/test split is done.

SOA2015 actually uses duration (in the form (Duration - 9)^(-n) for n = 1, 2, 3) as a predictor and claims that

This model also extends beyond the initial shock lapse all the way to duration 19

The dataset they provided doesn't really support this, as

> table(insurance::lapse_study$duration)

    10     11     12    13+    6-9 
 59730  29659  19233  35069 201936 

we'll do our best here and take 10, 11, and 12, and convert those to numeric

Describe process going from raw data from SOA to modeling dataset.

Describe how we tried to replicate the SOA2015 model.

This issue tracks writing the code to produce the plots.

We don't need to go into too much detail (out of scope) but we may want to show these and briefly describe what they are.

• Model vs. model performance
• Effects of predictors on response
• Prediction breakdown

Build the SOA2015 model as specified, to the best of our ability. This will be used as the state of the art baseline.

• SOA2015 (https://www.soa.org/research-reports/2015/lapse-2015-modeling-post-level/) This is the current "state of the art"
• AutoML with no feature engineering
• Maybe hand tune a neural net

• AutoML
• Neural net

We should do an initial training/testing split on a policy year boundary. Maybe 80/20. For the evaluation metric, weighted RMSE (with exposure as weights) should be reasonable.