If the models have parameters which are tuned purely based on data available up until a certain point in time, and can predict outcomes based on data that is observed after that point in time, then those models can be said to be good predictors.

For example, if a model is trained to calculate the average Earth surface temperature based on recorded CO2 levels up to the year 2000, and then it correctly predicts the average surface temperature of all subsequent years, with just the CO2 levels as input, that would be a very helpful model for deciding what would be a safe level of CO2 to have in the atmosphere if we want the Earth's temperature to have less than some particular amount of warming.