Azure AutoML time series model returns strange forecast

Question

I used Azure AutoML to train a time series forecasting model and selected the forecasting horizon to be 6. Each of our data row is one month, so we want to see the forecast for the following 6 months.

However, when feeding 2 rows of data to the model, it returns 2 figures, and when feeding 8 rows of data, it returns 8 figures. We expect that as we select the forecasting horizon to be 6, regardless of how many rows of data being fed into the model, it should returns 6 figures.

Could somebody explain why this happens and how to correct it? Thank you.

Answer 1

Hello,

I would suggest you double check how you are defining time_column_name. Thanks.

Attached the definition of this class for your reference,

fulldata: pandas.DataFrame           a time series dataset. Needs to contain X and y.
time_column_name: string             which column (must be in fulldata) is the time axis
target_column_name: string           which column (must be in fulldata) is to be forecast
forecast_origin: datetime type       the last time we (pretend to) have target values 
horizon: timedelta                   how far forward, in time units (not periods)
lookback: timedelta                  how far back does the model look?

Regards,
Yutong

Answer 2

I think it will forecast the number of rows you feed into the forecast function (feed datetime with empty column for output). So the behaviour you see is expected.

Ideally you feed in no more than your training forecast horizon. If you feed more it wil become increasingly inaccurate as it starts to use the previously forecast value as input for the "over the horizon" forecasts.