Skip to main content
Run StatsForecast distributedly on top of Ray.
StatsForecast works on top of Spark, Dask, and Ray through Fugue. StatsForecast will read the input DataFrame and use the corresponding engine. For example, if the input is a Ray Dataset, StatsForecast will use the existing Ray instance to run the forecast. A benchmark (with older syntax) can be found here where we forecasted one million timeseries in under half an hour.

​
Installation

As long as Ray is installed and configured, StatsForecast will be able to use it. If executing on a distributed Ray cluster, make use the statsforecast library is installed across all the workers.

​
StatsForecast on Pandas

Before running on Ray, it’s recommended to test on a smaller Pandas dataset to make sure everything is working. This example also helps show the small differences when using Ray. 
from statsforecast.core import StatsForecast
from statsforecast.models import AutoARIMA, AutoETS
from statsforecast.utils import generate_series

n_series = 4
horizon = 7

series = generate_series(n_series)

sf = StatsForecast(
    models=[AutoETS(season_length=7)],
    freq='D',
)
sf.forecast(df=series, h=horizon).head()
unique_iddsAutoETS
002000-08-105.261609
102000-08-116.196357
202000-08-120.282309
302000-08-131.264195
402000-08-142.262453

​
Executing on Ray

To run the forecasts distributed on Ray, just pass in a Ray Dataset instead. 
import ray
import logging

ray.init(logging_level=logging.ERROR)

series['unique_id'] = series['unique_id'].astype(str)
ctx = ray.data.context.DatasetContext.get_current()
ctx.use_streaming_executor = False
ray_series = ray.data.from_pandas(series).repartition(4)

sf.forecast(df=ray_series, h=horizon).take(5)