In this post we take a break from our recent events & leadership focus and get ‘down & dirty’ with technical detail. In this case the challenge for insurance businesses to build and refine their pricing models in an ever more competitive data-led battleground.
I’m not normally a fan of automated model building, given the importance of domain knowledge guiding variable selection etc. Done well, however, it can save time & money as long as you can still be assured of model quality. Of course there are times when the speed trade-off is worth it, even where rebuilding models more often can be an advantage.
Many years ago, when KXEN first launched on the market, my statisticians were sceptical of model quality and a mathematical optimisation technique that was too ‘black box’. In our trials it proved true that models built using their proprietary technique did degrade more quickly than those built using existing process. However, they could also be built in half the time. So there can be a trade-off (even though at that time we found the same benefit by using Portrait Miner more visual software).
Two things have persuaded me to share with you a new breakthrough. Firstly that it’s been made by a friend of mine, Tony Ward, who when working for me proved himself a very competent statistician & programmer. Secondly, that he has approached the problem of having a more open solution to automated/algorithmic model building by using ‘R’. This is both increasingly favoured by analytics and insight teams and makes the methodology transparent to any R user.
In the spirit of today’s content marketing, mobile/social & collaborative businesses, Tony is also giving away the R code to achieve this model building (and a white paper on his research, method and robust testing used). That really is a free offer that I can’t but share with our statistical and analytics leaders. So, here is Tony’s site for access to that paper & code (just sign-up at the bottom of the webpage, with no commitment or cost):
Algorithmic Pricing is a new framework for building Generalised Linear Models (GLMs) in an automated way, using Machine Learning. To download the white paper and starter code, please sign up at the bottom of this page. In this case study we benchmark the Algorithmic Pricing Approach against the Traditional GLM Approach, using a claims frequency dataset.
As Tony shares in the results published, this algorithmic method outperformed the traditional Actuarial method of building GLM models (in 9 out of 10 cases). Although the average improvement in R² of 0.04% might seem small (even though statistically significant), the projected benefits are not. Tony & his team estimate that a move to this method could improve an insurance company’s loss ratio by 1%, as well as increasing average premiums and contribution. That is a benefit well worth exploring.
Hope you found that diving into a technical opportunity an interesting change. The wider data science community often share useful R code snippets, it will be interesting to see how this practice manifests itself within commercial businesses & their customer insight teams.
How are you building pricing models? Any tips to share on automated model building?