Adding thief and mlp.thief forecasting functions in forecast.gts (i.e. fmethod = c("ets", "arima", "rw", "thief", "mlp.thief", "elm.thief"))?

Question

I have a hierarchical data. The data has the following levels (top to down):

1. Production facility
2. Industry
3. Costumer
4. Product group
5. SKU

I am forecasting using r library hts. To increase accuracy I would like to use thief library (also nnfor::mlp.thief and nnfor::elm.thief functions). I added these functions to forecast.hts() in the following way,

loopfn <- function(x, ...) {
    out <- list()
    if (is.null(FUN)) {
      if (fmethod == "ets") {
        models <- forecast::ets(x, lambda = lambda, ...)
        out$pfcasts <- forecast::forecast(models, h = h, PI = FALSE)$mean
      } else if (fmethod == "arima") {


        models <- forecast::auto.arima(x, lambda = lambda, xreg = xreg,
                             parallel = FALSE, ...)
        out$pfcasts <- forecast::forecast(models, h = h, xreg = newxreg)$mean


      } else if (fmethod == "rw") {
        models <- forecast::rwf(x, h = h, lambda = lambda, ...)
        out$pfcasts <- models$mean
      } else if (fmethod == "thief"){

        models <- thief::thief(x, h = h , usemodel = usemodel, ...)
        out$pfcasts <- models$mean
      }else if (fmethod == "mlp.thief"){

        models <- nnfor::mlp.thief(x, h = h , ...)
        out$pfcasts <- models$mean
      } else if (fmethod == "elm.thief"){

        models <- nnfor::elm.thief(x, h = h , ...)
        out$pfcasts <- models$mean
      }


    } else { # user defined function to produce point forecasts
      models <- FUN(x, ...)
      if (is.null(newxreg)) {
        out$pfcasts <- forecast(models, h = h)$mean
      } else {
        out$pfcasts <- forecast(models, h = h, xreg = newxreg)$mean
      }
    }
    if (keep.fitted) {
      out$fitted <- stats::fitted(models)
    }
    if (keep.resid) {
      out$resid <- stats::residuals(models)
    }
    return(out)
  }

Would there be any theoretical problem to do so? In fact, it increases forecast accuracy.

considering the following literature I do not see any problem though Hyndman et.al. 2017 and Hyndman et.al. 2011

Answer 1

I could use any function with the argument FUN = The function has to be forecast object.