In the popular R package forecast, there are 2 different types of interfaces:

• A direct interface for functions like forecast::thetaf doing fitting and inference simultaneously

nile.fcast <- forecast::thetaf(Nile)
plot(nile.fcast)

• An interface for fitting first and then forecasting, like forecast::ets, where you need to use, in addition forecast::forecast:

fit <- forecast::ets(USAccDeaths)
plot(forecast::forecast(fit))

In this post, I describe how to obtain probabilistic forecasts from R package forecast – and packages that follow a similar philosophy such as forecastHybrid, ahead, etc. –, by using a unified interface (ahead::genericforecast). Then, I present ahead::conformalize, a function that allows to obtain forecasts using the method described in Conformalized predictive simulations for univariate time series (more details can be found in these slides).

0 – Packages

utils::install.packages(c("remotes", "e1071", "forecast", "glmnet"))
remotes::install_github("Techtonique/ahead")

library(ahead)
library(forecast)

y <- fdeaths #AirPassengers #Nile #mdeaths #fdeaths #USAccDeaths
h <- 25L

1 – Generic forecaster (unified interface)

1 – 1 – Using default parameters

par(mfrow=c(2, 2))
plot(ahead::genericforecast(FUN=forecast::thetaf, y, h))
plot(ahead::genericforecast(FUN=forecast::meanf, y, h))
plot(ahead::genericforecast(FUN=forecast::rwf, y, h))
plot(ahead::genericforecast(FUN=forecast::ets, y, h))

par(mfrow=c(2, 2))
plot(ahead::genericforecast(FUN=forecast::tbats, y, h))
plot(ahead::genericforecast(FUN=HoltWinters, y, h))
plot(ahead::genericforecast(FUN=forecast::Arima, y, h))
plot(ahead::genericforecast(FUN=ahead::dynrmf, y, h))

1 – 2 – Using additional parameters

par(mfrow=c(2, 2))
plot(ahead::genericforecast(FUN=ahead::dynrmf, y=y, h=h, 
                            fit_func=e1071::svm, predict_func=predict))
plot(ahead::genericforecast(FUN=ahead::dynrmf, y=y, h=h, 
                            fit_func=glmnet::cv.glmnet, predict_func=predict))
plot(ahead::genericforecast(FUN=forecast::tbats, y=y, h=h, 
                            use.box.cox = TRUE, use.trend=FALSE))
plot(ahead::genericforecast(FUN=forecast::rwf, 
                            y=y, h=h, lambda=1.1))

2 – Conformal prediction

2 – 1 – Using default parameters

y <- USAccDeaths

par(mfrow=c(3, 2))
obj <- ahead::conformalize(FUN=forecast::thetaf, y, h); plot(obj)
obj <- ahead::conformalize(FUN=forecast::meanf, y, h); plot(obj)
obj <- ahead::conformalize(FUN=forecast::rwf, y, h); plot(obj)
obj <- ahead::conformalize(FUN=forecast::ets, y, h); plot(obj)

par(mfrow=c(2, 2))
obj <- ahead::conformalize(FUN=forecast::auto.arima, y, h); plot(obj)
obj <- ahead::conformalize(FUN=forecast::tbats, y, h); plot(obj)
obj <- ahead::conformalize(FUN=HoltWinters, y, h); plot(obj)
obj <- ahead::conformalize(FUN=forecast::Arima, y, h); plot(obj)

2 – 2 – Using additional parameters

y <- AirPassengers

par(mfrow=c(2, 2))
obj <- ahead::conformalize(FUN=forecast::thetaf, y, h); plot(obj)
obj <- ahead::conformalize(FUN=forecast::rwf, y=y, h=h, drift=TRUE); plot(obj)
obj <- ahead::conformalize(FUN=HoltWinters, y=y, h=h, seasonal = "mult"); plot(obj)
obj <- ahead::conformalize(FUN=ahead::dynrmf, y=y, h=h, fit_func=glmnet::cv.glmnet, predict_func=predict); plot(obj)

2 – 3 – Using other simulation methods (conformal prediction-based)

y <- fdeaths

par(mfrow=c(3, 2))
obj <- ahead::conformalize(FUN=forecast::thetaf, y=y, h=h, method="block-bootstrap"); plot(obj)
obj <- ahead::conformalize(FUN=forecast::rwf, y=y, h=h, drift=TRUE, method="bootstrap"); plot(obj)
obj <- ahead::conformalize(FUN=forecast::ets, y, h, method="kde"); plot(obj)
obj <- ahead::conformalize(FUN=forecast::tbats, y=y, h=h, method="surrogate"); plot(obj)
obj <- ahead::conformalize(FUN=HoltWinters, y=y, h=h, seasonal = "mult", method="block-bootstrap"); plot(obj)
obj <- ahead::conformalize(FUN=ahead::dynrmf, y=y, h=h, fit_func=glmnet::cv.glmnet, 
                           predict_func=predict, method="surrogate"); plot(obj)