Prophet

Prophet is an open-source time-series forecasting algorithm developed by Facebook. It is suitable for series with clear seasonality, holiday effects, and trend changes. Prophet is robust to missing values and outliers, automatically detects periodic patterns, and supports custom seasonality and holiday parameters. Unlike ARIMA, it does not require a fully stationary series and can model yearly, weekly, and daily seasonalities.

The model consists of:

• A trend component with linear or saturating growth and changepoints.
• A seasonality component fitted with Fourier series.
• A holiday component for custom holidays or events.

Parameters

Parameter	Description	Required
growth	Trend type: linear or logistic. Default: linear	No
yearly_seasonality	Enable yearly seasonality: true, false, or auto. Default: auto	No
weekly_seasonality	Enable weekly seasonality: true, false, or auto. Default: auto	No
daily_seasonality	Enable daily seasonality: true, false, or auto. Default: auto	No
changepoint_prior_scale	Trend-changepoint flexibility. Larger values allow more variation. Default: 0.05	No

Example

Forecast the daily passengers series with yearly and weekly seasonality and a changepoint prior scale of 0.1:

SELECT _frowts,
       FORECAST(passengers, "algo=prophet,growth=linear,yearly_seasonality=true,weekly_seasonality=true,changepoint_prior_scale=0.1")
FROM air;

The model result uses the following fields:

{
  "rows": fc_rows,
  "algo": "prophet",
  "growth": "linear",
  "yearly_seasonality": "auto",
  "weekly_seasonality": "auto",
  "daily_seasonality": "auto",
  "changepoint_prior_scale": 0.05,
  "mse": mse,
  "res": res
}

References

1. Prophet documentation
2. Forecasting at Scale (2017)