Greykite suitable for pure linear increasing series?

[akthammomani]

Hello

I'm working in some house price time series using Greykite but for some reason, the forecast I got is just a median price between upper and lower (ARIMA), so is this known issue with Greykite when we have a pure linear increasing series?

Thank you
Aktham Momani

[Reza1317]

Hi,

Could you provide more details?
What is the data frequency? daily?

Also can you provide the code you used to fir the model?

Thanks
Reza

[akthammomani]

Data frequency is monthly so I used freq="MS" as showing below:

Please let me know if additional information is needed.

metadata = MetadataParam(
time_col="ts",
value_col="y",
freq="MS"
)

forecaster = Forecaster()
result = forecaster.run_forecast_config(
df=df,
config=ForecastConfig(
model_template=ModelTemplateEnum.SILVERKITE.name,
forecast_horizon=12,
coverage=0.95,
metadata_param=metadata
)
)

snapshot of the data used:

ts datetime64[ns]
y float64

ts	                  y

0 2012-02-01 390000.0
1 2012-03-01 400000.0
2 2012-04-01 410000.0
3 2012-05-01 438000.0
4 2012-06-01 470000.0
... ... ...
106 2020-12-01 1030000.0
107 2021-01-01 1050000.0
108 2021-02-01 1085000.0
109 2021-03-01 1095000.0
110 2021-04-01 1125000.0

Regards,
Aktham

[sayanpatra]

@akthammomani Silverkite tries to automatically introduce some amount of seasonality to the model. However, it is probably redundant for your use case. You can try to:

1. Tune the seasonality (maybe even turn those off). Example here: https://github.com/linkedin/greykite/blob/master/docs/nbpages/tutorials/0100_forecast_tutorial.py (line 446)
2. Use a different template e.g. SILVERKITE_WEEKLY. It was tuned for weekly data, however might produce a better model than SILVERKITE. More details here: https://github.com/linkedin/greykite/blob/master/docs/pages/stepbystep/0200_choose_template.rst

[akthammomani]

Thank you for the recommendations.

Both of the suggestions, produced same initial forecast, as shown below when seasonality is off:

[sayanpatra]

@akthammomani Can you try using SILVERKITE_EMPTY template with C(month) as a regressor, ct1 as growth, and autoregression with "auto" or lag 1

[KaixuYang]

Close this issue as this thread has been inactive for many days.