auto.arima gives error wrong length for 'fixed'

[chintan-sd]

Hi, When using auto.arima with lambda parameter, I have a specific case where it is giving error wrong length for 'fixed'

I have posted details and an example to reproduce issue here: https://stats.stackexchange.com/questions/403990/auto-arima-throws-error-wrong-length-for-fixed

Thanks Chintan

[mitchelloharawild]

I'd be careful when using BoxCox.lambda() (or lambda="auto") for this series. The chosen lambda parameter seems very extreme for the series. Nonetheless, it seems to be something which can be fixed by changes in the forecast package.

---

If narma = 0 and nxreg=0, then fixed should be NULL.

MRE:

library(forecast)

t <- c(5354969,5714384,6993786,8134965,7291882,7855863,8260187,10187979,8472624,9317541,8667358,10245900,8140618,7501393,7592070,9683767,8327092,9042875,9835730,10828364,8200433,8516591,8290409,8406130,6587574,8174418)
myts <- ts(t, frequency = 4)

f <- forecast(auto.arima(myts, stepwise = FALSE, approximation = FALSE, lambda = "auto"), h = 2)
#> Error in stats::arima(x = x, order = order, seasonal = seasonal, include.mean = include.mean, : wrong length for 'fixed'

^{Created on 2019-04-20 by the reprex package (v0.2.1)}

[mitchelloharawild]

@robjhyndman I don't fully understand the logic here: https://github.com/robjhyndman/forecast/blob/master/R/newarima2.R#L290 What is this trying to achieve? Add a drift term to the model? It conflicts with the condition in https://github.com/robjhyndman/forecast/blob/master/R/arima.R#L715

Resulting in no drift term being included in the model.

[robjhyndman]

The first is inside an if statement for constant data after differencing. If the data are constant, and d+D=1, then we want to add a drift term so that the forecasts have the same property. If d+D=0, we want to add an intercept. If d+D > 1, it should probably not include the drift. So yes, this could be improved.

The second is inside an if statement which tests if the user wants a constant in the model (not the same as constant data). In that case, the constant is equivalent to a drift term only if d+D=1.

[mitchelloharawild]

So I guess the issue is that the first condition in newarima2.R#L290 is being satisfied as D=1 and d=1, when the condition originally intended to capture D=1 and d + D = 1? Otherwise the constant is being requested from the model with a fixed coefficient in newarima2.R#L290, but the regressor is never being added to the model as d + D > 1 in arima.R#L715