Hi Team, I am getting a weird issue with TimeGPT. I have a historical series with only 1 data point with a value of 1. When I forecast this, I get a forecast of 236. This is part of a bigger dataset, but I have isolated this to a single series to debug. Reproducible example below. Does anyone know what is going on here and how to fix this? Thanks!

ID_COL = "unique_id"
TIME_COL = "date"
TARGET_COL = "qty"

new = pd.DataFrame({
    ID_COL: ['270*28'],
    TIME_COL: pd.date_range(start='2022-10-31', periods=1, freq=freq),
    TARGET_COL: [1]    
})

forecast_new = nixtla_client.forecast(
    df=new,
    **{
        'h': 52,
        'freq': 'W-MON',
        'model': 'timegpt-1-long-horizon',
        # 'date_features': ['week', 'month', 'year'],
        'id_col': 'unique_id',
        'time_col': 'date',
        'target_col': 'qty'
    }
)

And that's why I said you should add it to the list ofnsupported models

I think it's probably due to the limited amount of datapoints, but of course we should be able to easily do better here.... (just forecast the same value)

The problem is that there are too many series like this and it is really inflating the top level forecasts (aggregated)

nixtlaclient.supported_models.append("name of model")

🙂

it's just a plain list, you can add models to it

Much better

(I'm on my phone, so the syntax might not be exactly correct)

No problem that works

Well the easy way in this case is then to just switch to 1.5, but it's not like we guarantee it doesn't happen in 1.5

Yes, also, we have limited time for POC now. So this will require more work from them.