ARIMA with a period of less than 2 years

[LucasGBegiato]

Is it possible to make an ARIMA forecasting with a dataset containing less than 2 years in observations?

------------------ Code ------------------------------------------------

database <- read.csv("STcomFAIXAetaria.csv", header=TRUE, sep=";")

data1.ts <- ts(database[,"X0.a.9.anos"], frequency = 12, start = c(2020,2))
data1.ts

fit.ARIMA <- auto.arima(data1.ts, seasonal = FALSE)
print(summary(fit.ARIMA))
checkresiduals(fit.ARIMA)

fcast <- forecast(fit.ARIMA, h=2)
plot(fcast)

--------------------------------------------------------------------------

It doesn't feel quite right. Am I doing something wrong? To make an arima forecast do I need two full periods?

[startz]

Is it possible? Sure. But ARIMA estimation is designed to work in relatively large samples. You're right to be nervous.

[LucasGBegiato]

Do you know how can I do it? My results don't show me anything. Am I right?

I would like to know how I could model this data by an ARIMA, since it has a period of less than two years.

Is there any package for this?

[startz]

What are you looking at when you say the "results don't show me anything?"

An ARIMA model has autoregressive, moving average, and differencing parameters. You are letting an algorithm pick the parameters. It is choosing a first difference and nothing else. With such a small sample, you might be better off trying some simple models manually using the arima() function.

[LucasGBegiato]

I mean that when I do auto.arima, it always results in ARIMA(0,1,0), that is, there is only one differentiation and nothing else, even when I change to different datasets.

Alright, I'm going to try this. And to see if it's good, is it enough to check the Ljung-Box test?

[mitchelloharawild]

auto.arima() selects a model by minimising AICc - this information criterion scores a model's performance with consideration of the number of observations and model parameters. When you don't have a lot of data, it will choose simpler (but still good) models with less parameters. This is why it is consistently choosing ARIMA(0,1,0), which isn't necessarily a bad model when you have very few observations to work with.