midasr (version 0.9)

average_forecast: Average forecasts of MIDAS models

Description

Average MIDAS model forecasts using specified weighting scheme. Produce in-sample and out-of-sample accuracy measures.

Usage

average_forecast(
  modlist,
  data,
  insample,
  outsample,
  type = c("fixed", "recursive", "rolling"),
  fweights = c("EW", "BICW", "MSFE", "DMSFE"),
  measures = c("MSE", "MAPE", "MASE"),
  show_progress = TRUE
)

Value

a list containing forecasts and tables of accuracy measures

Arguments

modlist

a list of midas_r objects

data

a list with mixed frequency data

insample

the low frequency indexes for in-sample data

outsample

the low frequency indexes for out-of-sample data

type

a string indicating which type of forecast to use.

fweights

names of weighting schemes

measures

names of accuracy measures

show_progress

logical, TRUE to show progress bar, FALSE for silent evaluation

Author

Virmantas Kvedaras, Vaidotas Zemlys

Details

Given the data, split it to in-sample and out-of-sample data. Then given the list of models, reestimate each model with in-sample data and produce out-of-sample forecast. Given the forecasts average them with the specified weighting scheme. Then calculate the accuracy measures for individual and average forecasts.

The forecasts can be produced in 3 ways. The "fixed" forecast uses model estimated with in-sample data. The "rolling" forecast reestimates model each time by increasing the in-sample by one low frequency observation and dropping the first low frequency observation. These reestimated models then are used to produce out-of-sample forecasts. The "recursive" forecast differs from "rolling" that it does not drop observations from the beginning of data.

Examples

set.seed(1001)
## Number of low-frequency observations
n<-250
## Linear trend and higher-frequency explanatory variables (e.g. quarterly and monthly)
trend<-c(1:n)
x<-rnorm(4*n)
z<-rnorm(12*n)
## Exponential Almon polynomial constraint-consistent coefficients
fn.x <- nealmon(p=c(1,-0.5),d=8)
fn.z <- nealmon(p=c(2,0.5,-0.1),d=17)
## Simulated low-frequency series (e.g. yearly)
y<-2+0.1*trend+mls(x,0:7,4)%*%fn.x+mls(z,0:16,12)%*%fn.z+rnorm(n)
mod1 <- midas_r(y ~ trend + mls(x, 4:14, 4, nealmon) + mls(z, 12:22, 12, nealmon),
                start=list(x=c(10,1,-0.1),z=c(2,-0.1)))
mod2 <- midas_r(y ~ trend + mls(x, 4:20, 4, nealmon) + mls(z, 12:25, 12, nealmon),
                start=list(x=c(10,1,-0.1),z=c(2,-0.1)))

##Calculate average forecasts
avgf <- average_forecast(list(mod1,mod2),
                        data=list(y=y,x=x,z=z,trend=trend),
                        insample=1:200,outsample=201:250,
                        type="fixed",
                        measures=c("MSE","MAPE","MASE"),
                        fweights=c("EW","BICW","MSFE","DMSFE"))