Choose a Model
Greykite offers a few forecasting models: Silverkite, Prophet, and ARIMA. This page provides an overview.
SILVERKITE |
PROPHET |
ARIMA |
|
---|---|---|---|
speed |
fast |
slow |
fast |
forecast accuracy (default) |
decent |
decent |
decent |
forecast accuracy (customized) |
very good |
good |
good |
interpretability |
good |
good |
decent |
ease of use |
good |
good |
very good |
API |
|
similar to |
similar to |
Like Prophet, Silverkite includes intepretable terms for growth, seasonality, holidays, trend changepoints, and regressors. Silverkite also supports autoregression, seasonality changepoints, easy-to-use interaction terms, quantile loss, and custom fit algorithms. This makes Silverkite flexible to capture different time series patterns.
SILVERKITE |
PROPHET |
|
---|---|---|
automatic defaults |
yes
|
yes
|
growth |
linear, sqrt, quadratic,
any combination, custom
|
linear, logistic
|
seasonality |
daily, weekly, monthly,
quarterly, yearly
|
daily, weekly, yearly, custom
|
holidays |
specify names or countries,
with window; or custom events
|
specify countries, with window
|
regressors |
yes
|
yes
|
trend changepoints |
yes
|
yes
|
seasonality changepoints |
yes
|
no
|
autoregression |
yes
|
limited
(via regressors)
|
interaction terms |
yes
(via model formula or regressors)
|
limited
(via regressors)
|
loss function |
MSE, Quantile loss
|
MSE
|
fit algorithm |
custom
(ridge, quantile regression, etc.)
|
fixed
(Bayesian formulation)
|
Note
When to use the Silverkite model?
If both speed and interpretability are important
If you need flexible tuning options to achieve high accuracy
If you need to forecast a quantile
Note
When to use the Prophet model?
If you need logistic growth with changing capacity over time
If speed is not as important
Note
When to use the ARIMA model?
If you want to try a classic algorithm that is different from the other two
If you want to quickly establish an accuracy baseline to assess forecast difficulty
If interpretability is not as important