Rolling window option in tsCV() with xreg

[mkiang]

Hi, I may be a bit confused here, but I'm having trouble setting the window option in tsCV() while also using xreg. When not using the xreg, the window seems to scan the data correctly:

far2 <- function(x, h){forecast(Arima(x, order=c(2,0,0)), h=h)}
e <- tsCV(lynx, far2, h=1, window=30)

fit1   <- Arima(lynx[1:30], order = c(2, 0, 0))
fcast1 <- forecast(fit1, h = 1)
lynx[31] - fcast1$mean[1]; e[30]

# [1] -11.41075
# [1] -11.41075

fit2   <- Arima(lynx[2:31], order = c(2, 0, 0))
fcast2 <- forecast(fit2, h = 1)
lynx[32] - fcast1$mean[1]; e[31]

# [1] -578.658
# [1] -578.658

When using the same idea as above, but specifying an xreg, it does not work. Any tips on how to write my forecastfunction such that I can use a rolling window and xreg? It works for the first fit (i.e., before the window starts rolling), but not for other fits.

fc <- function(y, h, xreg)
{
    X <- xreg[1:length(y),]
    if(NROW(xreg) < length(y) + h)
        stop("Not enough xreg data for forecasting")
    newX <- xreg[length(y)+(1:h), ]
    fit <- Arima(y, xreg=X, order = c(1, 0, 0))
    forecast(fit, xreg=newX)
}

set.seed(1)
y <- ts(rnorm(100))
x <- matrix(ts(rnorm(100)),ncol=1)
e <- tsCV(y, fc, xreg=x, window = 20)

fit1   <- Arima(y[1:20], xreg = x[1:20, ], order = c(1, 0, 0))
fcast1 <- forecast(fit1, h = 1, xreg = x[21, ])
y[21] - fcast1$mean[1]; e[20]
# [1] 0.7328816
# [1] 0.7328816

fit2   <- Arima(y[2:21], xreg = x[2:21, ], order = c(1, 0, 0))
fcast2 <- forecast(fit2, h = 1, xreg = x[22, ])
y[22] - fcast2$mean[1]; e[21]
# [1] 0.7791716
# [1] 0.5096576

Thanks for any help!

[robjhyndman]

Should be ok now. Fixed in 904028f

[robjhyndman]

@mitchelloharawild might be able to make the code a bit neater though.

[anafotina]

When reproducing the example above, the values between tsCV and manually calculated errors are still different starting from the second step.