Comment by quantumofalpha

Why not both? Linear models for absolute lowest tick-to-trade latency doesn't preclude you from using fancier stuff at earlier modelling steps. Final linear models you ship to fpga can be mere distillations/triggers