I have a time series that looks like this t value 1 12 2 12 3 0 4 0 5 0 6 0 7 0 I expect acf1 to equal 0.443, but instead the accuracy function produces null. The code is as following:
df = data.frame("t" = 1:7, "value" = c(12, 12, 0, 0, 0, 0, 0))
tsb = df %>%
as_tsibble(index = t)
md = tsb %>% model(arima = ARIMA(value ~ PDQ(period = 4), stepwise = F))
fc = md %>% forecast(h = 4)
accuracy(fc, tsb)
Why is this happening?
The ACF1 column from accuracy() is the first auto-correlation of the residuals. The ACF1 of 0.443 that you expect is the first auto-correlation of your data, which can be obtained with:
library(feasts)
#> Loading required package: fabletools
df = data.frame("t" = 1:7, "value" = c(12, 12, 0, 0, 0, 0, 0))
tsb = df %>%
as_tsibble(index = t)
tsb %>% ACF(lag_max = 1)
#> Response variable not specified, automatically selected `var = value`
#> # A tsibble: 1 x 2 [1]
#> lag acf
#> <lag> <dbl>
#> 1 1 0.443
Created on 2020-08-13 by the reprex package (v0.3.0)
The second problem with your usage is that accuracy() for forecasts requires future data to compute the forecast errors. The forecasts in fc do not match the times provided by tsb, and so no forecast errors can be computed.
library(tsibble)
library(dplyr)
library(fable)
md = tsb %>% model(arima = ARIMA(value ~ PDQ(period = 4), stepwise = F))
fc = md %>% forecast(h = 4)
# Make up some future data for evaluating forecast accuracy
tsb_future <- new_data(tsb, 4) %>% mutate(value = rnorm(4))
# Compute the accuracy of the forecasts against the tsb_future scenario
accuracy(fc, tsb_future)
#> # A tibble: 1 x 9
#> .model .type ME RMSE MAE MPE MAPE MASE ACF1
#> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 arima Test -0.779 1.09 0.975 100 100 NaN -0.0478
Created on 2020-08-13 by the reprex package (v0.3.0)
