Choose a Model

Greykite offers two forecasting models: the Prophet model and the Silverkite model. This page explains your options.

high-level comparison

	PROPHET	SILVERKITE
speed	slower	faster
forecast accuracy (default)	good	good
forecast accuracy (customized)	limited	high
prediction interval accuracy	TBD	TBD
interpretability	good (additive model)	good (additive model)
ease of use	good	good
API	similar to sklearn	uses sklearn
fit	Bayesian	ridge, elastic net, boosted trees, etc.

Both models have the similar customization options. Differences are bolded below.

customization options

	PROPHET	SILVERKITE
automatic defaults	yes	yes
growth	linear, logistic	linear, sqrt, quadratic, any combination, custom
seasonality	daily, weekly, yearly, custom	daily, weekly, monthly, quarterly, yearly
holidays	specify countries, with window	specify by name or country, with window; or custom events
trend changepoints	yes	yes
seasonality changepoints	no	yes
regressors	yes	yes
autoregression	limited, via regressors	full support, coming soon
interaction terms	build it yourself (regressor)	model formula terms, or as regressor
extras	prior scale (bayesian)	fitting algorithm
loss function	MSE	MSE, Quantile loss (with gradient_boosting fitting algorithm)
prediction intervals	yes	yes

Note

When to use the Prophet model?

• If it works better for your dataset

• If you like Bayesian models

• If you need logistic growth with changing capacity over time

Note

When to use the Silverkite model?

• If it works better for your dataset (e.g. b/c of custom growth, interaction terms, seasonality changepoints).

• If speed is important.

• If you want to forecast a quantile, not the mean.

Note

We use Prophet 0.5 (Prophet documentation.)