Transparency in GBMs

GBM models are often called "Black Box" when compared to GLMs because they are thought to not be transparent with more going on under the bonnet that isn't clear. The implication is that GLMs are either transparent or far more transparent by comparison. But I'm not sure that this is true.

Imagine a pricing structure that uses the same data in the same groupings, with one fed into a GLM and another into a GBM. The input data is the same, so the difference comes down to how that data is used. A GLM will use a set number of rating tables (some with interactions) whilst a GBM will use a tree structure. Both work by looking at the data inputs, comparing to the next decision point (table or node), and then multiplying all the relativities together.

The primary difference is that a GLM is not order sensitive but a GBM is (you could even 'reduce' a GBM into a table format, it would just be huge, repetitive, and inefficient). The basic mechanism is the same, except that a GLM will use each table once and can be calculated in any order, whilst to a GBM the order matters because each branch of the trees have to be done in a certain order.

Once you add this order sensitivity then the number of output options increases a lot. So the issue here is not that GLMs are inherently more transparent, but that their relative simplicity allows us to understand them.

But is that true? GLMs feel simpler because there are nice 1-way tables that you can look at and see the relativity for. But even a very simple rating structure of, say, 15 rating tables with an average of 8 rows per table would give us a 8^15 different answers, or 35,184,372,088,832 different answers (over 4,000 different answers for each person alive today). Can you really hold all of that in your head at the same time? Do you really understand what is going on once you start to multiply tables together?

On the reverse side you can build a GBM, score it against a representative batch of data, and then build a GLM for that dataset to get some table based relativities. It doesn't describe the full output, but then (GLM or GBM) no human can understand the full output anyway. But it does give us that sense of where prices are higher and lower that GLM tables provide. And for transparency isn't that what we are really after?

If you want to upgrade your GLMs but are concerned that GBMs don't give you the transparency that you want then get in contact today and we can have a chat about your options.


Continue reading