To measure the deforestation reduced by a policy, we need to compare deforestation rates under a policy with deforestation rates in the absence of policy. Unfortunately the deforestation rate in the absence of a policy, or reference rate, is ex ante difficult to forecast and ex post impossible to observe. This means that reference rates will be set with error and we will not know how large the error will be. The challenging nature of setting reference rates is reflected in the number of proposals for reference rate design. In this paper I show how these proposals ignore forecast error. As a consequence, these proposals have basic structural weaknesses that increase the costs of reduced deforestation policy. I propose that a criteria for reference rates is to minimise the cost of forecast error. These ideas are illustrated with a cross country dataset on agricultural expansion. I show that the best forecasting model differs by country and that a country’s best forecasting model can be very simple.