Problem

We run the MLR model on sequence counts aggregated at 14-day intervals, the step size of the model’s time scale is 14 days instead of 1 day.

When we forecast future frequencies to L steps into the future, those should be on the same time scale of 14-day steps, but the forecast dates exported to the model JSON occur at 1-day steps.

Proposed solution

To get the appropriate forecast dates from evofr, we could modify the forecast_dates function to infer the duration of time between steps from the observed dates passed as the first input to that function and use that duration instead of 1 day to generate the forecast dates