search

mages / ChainLadder

Claims reserving models in R
https://mages.github.io/ChainLadder/
75 stars 62 forks source link

User defined development factor in BootChainLadder #50

Open ludovictheate opened 6 years ago

ludovictheate commented 6 years ago

Hi,

I think it would be very useful to add an option in the BootChainLadder function in order to allow the user to force development factors other than those stemming from a pure application of the chainladder method. Indeed the calibration of development factors encompasses a certain level of expert judgement which can give rise to user defined development factors. The bootstrap should then be based on those DF and not on the canonical ones.

Is this something possible ? Thanks.

mages commented 6 years ago

This might be possible, but I am not sure this would make sense from a statistical point of view. It sounds too much like a fudge.

ludovictheate commented 6 years ago

Hi Markus.

Apologies but I don't see why this would be a fudge. It is market best-practice (and also a main feature in many of the well known reserving software) to perform adjustments on the development factors (removing outliers, reducing the calibration history,...). Performing a bootstrap using the canonical factors when the "best-estimate" is computed using adjusted factors doesn't make a lot of sense.

mages commented 6 years ago

Hi Ludovic, I am aware of the practice of selecting 'suitable' factors. However, it appears to me more based expert judgment, i.e. which data to include or exclude, rather than statistical, and therefore forcing a model to work with a given data set, instead of selecting a suitable model for the data at hand. Anyhow, that's more a philosophical point and what I would call 'fudging'. Yet, I am happy to support you, if you would like to look into the implementation of your idea.

trinostics commented 6 years ago

The idea of using data to estimate the variability of a set of judgmentally selected factors is basis of the paper I wrote with Bardis and Majidi that extends the Mack/Murphy method in that situation. The paper is here http://www.variancejournal.org/issues/06-02/143.pdf, Bardis' presentation at a CAS meeting is here: https://www.casact.org/education/spring/2013/handouts/Paper_2331_handout_970_0.pdf I blogged about how to implement the technique with ChainLadder back in 2014 ( http://trinostics.blogspot.com/2013/07/implementing-clfm-with-chainladder.html )

Ludovic appears to want to explore that possibility with the England's Bootstrap method. We got pushback from some folks too. One should expect to see a higher standard error of the prediction to the extent the selections don't agree with the canonical factors, which is exactly what we found. If that is the only dataset available for one's estimates, then it is nice to have an algorithm that is able to respond with a scientifically based result. Eventually we were able to convince our reviewers of the value of our thesis.

On Tue, Jun 19, 2018 at 9:15 AM, Markus Gesmann notifications@github.com wrote:

Hi Ludovic, I am aware of the practice of selecting 'suitable' factors. However, it appears to me more based expert judgment, i.e. which data to include or exclude, rather than statistical, and therefore forcing a model to work with a given data set, instead of selecting a suitable model for the data at hand. Anyhow, that's more a philosophical point and what I would call 'fudging'. Yet, I am happy to support you, if you would like to look into the implementation of your idea.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mages/ChainLadder/issues/50#issuecomment-398457931, or mute the thread https://github.com/notifications/unsubscribe-auth/AGKcB5ae07M_oCjuv-Qv-N3fByEXTZqXks5t-SO-gaJpZM4UtYrf .