DevelopmentML estimator

[jbogaardt]

With sparse triangle support, index and column values, and tight integration with sklearn already available, it shouldn't take too much effort to blend ML regression techniques with the convenient reserving properties of the chainladder package.

Refer to the ideas presented here -
https://www.casact.org/education/clrs/2016/presentations/AR-1_1.pdf

Something like:

cl.DevelopmentML(
   estimator=sklearn.ensemble.RandomForestRegressor()
).fit(cl.load_sample('prism')).ldf_

We'd need some way of providing control over feature selection and feature engineering, but otherwise seems doable.

[jbogaardt]

GLM is such an important framework in actuarial science it made sense to develop a separate TweedieGLM development esimtator just for that. Under the hood it uses the DevelopmentML estimator. The TweedieGLM is similar to the R ChainLadder glmReserve, but where we differ is:

1. I haven't fleshed out the prediction error, and they are thorough there
2. Instead I focused on flexibility in adding covariates (jurisdiction, cause of loss, limits, report lag, etc) to the model. In theory it could be used for policy/claim level loss reserving since the sparse backend allows for high throughput of data in chainladder. Flexibility come by way of the patsy package, but its a reasonably small dependency for a lot of flexibility.

Pre-release docs:
https://chainladder-python.readthedocs.io/en/latest/modules/development.html#tweedieglm

[andrejakobsen]

Are there plans to continue developing this feature? I could try to contribute, if I find the time.

[jbogaardt]

Hello @andrejakobsen, welcome! Yes, I would like to see this developed out more. Time is definitely the most constraining factor. If you have ideas, I would like to get them on board.