To be more specific, loss reserving models in actuarial science, such as the chain ladder method, can be expressed as GLMs. I have developed a predictive model using neural nets which takes into account some aspects of the insured (it is an individual risk model). Can the output of this model be safely used as an input to the insurance company's existing loss reserve model?


Theoretically there is no problem. I've seen tree models put as predictors in logistic models. NN as input into a GLM moodel makes sense. The ultimate decition should be made based on the predictability of the NN.

You have to mind a few issues:

  1. model mantenance and deployment. The NN model would probably have more parameters than your vanilla GLM deployment might be more involved. You could live it frozen and never update it, and let the GLM model use the NN score as long as it adds to the GLM prediction.

  2. interpretability. There might be managers whose acountability is to review risk, and they may not be comfortable with a NN. In that case feeding the NN results into a GLM might make it more acceptable, much as credit scores are used in some risk models.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.