Risk adjustment systems, that reallocate funds among competing health insurers, often use risk adjustors that are based on utilization. The level of utilization that triggers an adjustor - the utilization threshold - is frequently chosen implicitly and uniformly. I empirically study utilization thresholds in the setting of the U.S. Marketplaces and demonstrate how an explicit choice of such thresholds, tailored to each adjustor, may improve the prediction fit of the risk adjustment system and may decrease the incentives to game it. Using simulations, I find that a single alternative threshold may improve the prediction fit in some disease groups, by up to 14\%. A choice of multiple utilization thresholds, guided by a regression tree algorithm, may improve fit furthermore while taking into account the effect on gaming incentives.