Sometimes I've seen some rare cases, where the nnetar-forecast produces unrealistic low forecasts. Since it is a neuronal net you probably can't do much about it, but when you don't have an example you certainly can't do something about it. So I'm providing an example anyways, maybe it helps.
x <- timetk::tk_ts(c(
0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000,
0.000, 0.000, 0.000, 4344.544, 12642.858, 12642.858, 12068.183,
18678.098, 13969.492, 3034.287, 14482.585, 15928.499, 20226.667,
21239.201, 0.000
),
start = 1,
frequency = 12
)
fit <- forecast::nnetar(x)
forecast::autoplot(forecast::forecast(fit, h = 24))
It would be lovely, if you found a way to improve the nnetar-forecast for this kind of timeseries.
Sometimes I've seen some rare cases, where the nnetar-forecast produces unrealistic low forecasts. Since it is a neuronal net you probably can't do much about it, but when you don't have an example you certainly can't do something about it. So I'm providing an example anyways, maybe it helps.
It would be lovely, if you found a way to improve the nnetar-forecast for this kind of timeseries.