Say one has a predictive online model for market making or HFT (or just for anything strictly latency-constrained). In my specific example, I start with a Gaussian distribution over the "true value" of something I'm market making with and I have a method for updating a p.d.f. over it based on a classification of "informed" order flow. It is somewhat based on the Glosten-Milgrom Model if people are familiar with that, however really I am curious about a broader approach people have for dealing with this problem versus just my example. Put simply, you want to update your model enough s.t., well, your model is being utilized—but since market making is also a fast-paced game you need to avoid being run over.
Am I correct that this is more latency-constrained for the most part than anything else? The first step would be testing how computationally expensive it is to update the distribution obviously, but I'm curious about any preliminary insight anyone can provide. Thanks!
Additionally, when it comes to rigorously measuring and testing this: might there be any somewhat non-trivial metrics to look into?