Hi 🙂 I'm trying to better understand how the "update" functionality works. I have an hourly forecasting model trained up until 2024-01-01 00:00, with a maximum forecast horizon of 96 hours and using lagged features. When I update the model with new data (using the columns:

time_col

,

id_col

, and

target_col

, which in my case are

dt_hour_local

,

segment

, and

total_consumption_kwh

), for example with the following record

[2024-03-10 00:00, segmentA, 1.000]

I’m able to forecast starting from this updated time step. My questions are: 1. Why is this possible even though there’s a significant gap between the last training point (2024-01-01) and the update point (2024-03-10)? 2. What happens to the missing data between these two dates (2024-01-01 and 2024-03-10)? Is the model performing some kind of internal imputation or rolling forecast for those missing values before forecasting the future? Here is a depiction of what happens if you are not cautious about the model update: Blue = true Orange = prediction result when forecasting with "weird" model update - gap between last date seen by the model (model.ts.last_dates[0]) and the forecast start Green = result of correct update

Hey. What happens is that the new value (

1.0

) is appended to the historic targets and the last observed date is set to

2024-03-10

, so the forecasts start from that point. That's wrong though, since the lag features won't be computed correctly, there's currently no validation on the dates that are provided (issue)

Ah I see, thank you very much!