In the paper: https://robjhyndman.com/papers/temporalhierarchies.pdf , formula (1) seems to indicate the only temporal aggregation Thief allows is summation. Is it possible to include other aggregation method, such as quantile cut?

Is there a way to show the forecast interval with autoplot? It seems the "PI=TRUE" doesn't work.

Hi,
Can this be used for data sampled every five seconds for two hours, for forecasting for next 5 minutes?

I'm trying to optimize which methods I use at each temporal aggregate. In order to do this I must use reconcilethief(). I can get this function to work with other data sets such as AirPassengers, but for some reason not this. I receiver the error: "Error: length(n) == 1L is not TRUE". If you want to see it work, change tsaggregates(train) to tsaggregates(AirPassengers). Consider:

tsdata <- c(5669,6845,7320,7443,8031,8394,9321,8904,7982,7981,
7896,7733,8618,7013,8088,7069,6380,7171,7091,6491,
6184,6881,5480,7068,7004,6001,6890,6192,5904,6676,
6444,6181,5842,5861,5511,6941,7838,7185,7034,6446,
5920,6221,5391,7064,5687,5385,5929,7345,7157,6681,
8043,5570,6590,7365,6806)
train <- ts(tsdata, start=c(2013,1), f=12)
seriesagg <- tsaggregates(train)

model1 <- forecast(tslm(seriesagg$Monthly ~ season), h=48)
model2 <- forecast(snaive(seriesagg$'2-Monthly', h=24), h=24)
model3 <- thief(seriesagg$Quarterly, h=16)
model4 <- forecast(seriesagg$'4-Monthly', h=12)
model5 <- forecast(seriesagg$Biannual, h=8)
model6 <- forecast(seriesagg$Annual, h=4)

fc <- list()
fc[[1]] <- model1
fc[[2]] <- model2
fc[[3]] <- model3
fc[[4]] <- model4
fc[[5]] <- model5
fc[[6]] <- model6
reconciled <- reconcilethief(fc)

I am running thief (version 0.3) with 'usemodel' set to 'theta' and getting NaN as forecast mean.
here is 2 examples with quarterly series.
ts1 <- c(1820,1640,1890,2340,2480,2910,2800,2960,3150,3290,3100,2290,2070,2240,2030,2270,2200,2350,2580,2740,3050,3060,3280)

ts2<-c(2993,954,2072,2078,911,1841,3196,2648,1486,1261,1516,1314,1029,1948,1545,1723,1723,1925,2858,2624,2066,2354,2306)

Thank you.
Arsa

I am attempting to use the forecast package baggedETS with the thief package. I'm running into an error when attempting to do so, but when I use the same function on each aggregate level it works fine. Consider:

AirPassengers2 <- window(AirPassengers, start=c(1956,1))
model1 <- thief(AirPassengers2, h=41) # works

bagets <- function(y, h,...){forecast(baggedETS(y), h)} # attempting to use bagged ETS model

# Does not work
model2 <- thief(AirPassengers2, forecastfunction=bagets, h=41)

# individually they work
agg <- tsaggregates(AirPassengers2)
t1 <- bagets(agg[[1]],h=36)
t2 <- bagets(agg[[2]],h=18)
t3 <- bagets(agg[[3]],h=12)
t4 <- bagets(agg[[4]],h=9)
t5 <- bagets(agg[[5]],h=6)
t6 <- bagets(agg[[6]],h=4) # works now but generates warning from HoltWintersZZ

Previously I figured out baggedETS wouldn't work on < 8 observations. This was fixed (thank you!), but a new error now occurs:
Error in [.default(fc$lower, , 1) : incorrect number of dimensions

I apologize if this is a forecast package error, I'm not familiar enough with the source code to know where the root cause lies.

Hi Rob,

Thanks for another awesome package, I was wondering if there is a way to use regressors while aggregating temporal hierarchical forecasts ?

Thanks,
Sagar

Currently the thief() function cannot handle yearly data, presumably because there is no hierarchy to estimate. Instead of it erroring out, could it just trigger a warning and then default to the forecast function used?

data(M3)
thief(M3[[1]]$x, h = 1)

Error in thief(M3[[1]]$x, h = 1) :
Seasonal period (m) must be greater than 1

When running hts forecast would get Error in -.default(y, fits) : non-conformable arrays
is the thief function not returning fitted values in an easily accessible form resulting in the described error?

Code:
forecast(object = x,
h = 12,
method = "comb",
keep.fitted = TRUE,
weights = "mint",
positive = TRUE,
FUN = function(i) thief(i,m=12))

Can you provide a example of how to make reconcile thief work with comb equals shr?

Is it possible to run thief with forecastfunction equal to user-defined stlf function while stlf function has its own user-defined forecastfunction which is tbats?

I get below error in stlf function when running thief on it.
Error in stlf(x, h = h) : y is not a seasonal ts object

Thanks,

great code
but do you know something similar in python?

Will this 'thief' package will be installed in r studio. I tried this

install.packages('thief', dependencies = TRUE)

in r studio but i got installation error.

Thanks

At the end of this post: https://robjhyndman.com/hyndsight/thief/ , the reconciled forecasts seem to have prediction intervals. However, I can't find an argument in thief functions that enable prediction intervals.

Hi there,
Not sure this is the proper channel...complete newbie here. sorry.

With input

install.packages("devtools") devtools::install_github("robjhyndman/thief") library(thief)

RStudio gives:

install.packages("devtools")

devtools::install_github("robjhyndman/thief")
Downloading GitHub repo robjhyndman/thief@master
from URL https://api.github.com/repos/robjhyndman/thief/zipball/master
Installing thief
"C:/PROGRA1/R/R-331.1/bin/x64/R" --no-site-file --no-environ --no-save --no-restore --quiet
CMD INSTALL
"C:/Users/dh/AppData/Local/Temp/Rtmpaidvgs/devtools21b419b2650a/robjhyndman-thief-9302849"
--library="C:/Users/dh/Documents/R/win-library/3.3" --install-tests
ERROR: dependency 'TStools' is not available for package 'thief'

• removing 'C:/Users/dh/Documents/R/win-library/3.3/thief'
  Error: Command failed (1)

What should I try next?

I'm getting an odd behavior when trying to use the thief() function. What is interesting is that the function runs fine with my time series but when I add one observation to the time-series I get the error message:

"Error in [.default(fc$lower, , 1) : incorrect number of dimensions"

The reproducible bug (MRE), with data, can be found in

MRE URL:
https://gist.github.com/pgg1309/3ef3d0c5baf459af0262ffd08275ceee

Many thanks.
Paulo