# Interpreting and forecasting using ARIMA(0,0,0) or ARIMA(0,1,0) models

#1

Hi All,

I have time series data with 33 data points, however 29th data point has a sudden peak and when confirmed with business this peak is due to some changes and not expected anytime in future. So I decided to predict the 29th month using earlier 28 months data and then use the series for future predictions. I did initial analysis for stationarity and first order difference works in this case but the auto.arima gives ARIMA(0,0,0) model which is nothing but the white noise. Also, when I applied auto.arima on original series with all the obs it gives ARIMA(0,0,0)(0,1,0)[12].

My question is - how to get rid of the peak in 29th month? Any comments on ARIMA(0,0,0) models?

#2

Check for Seasonal random walk model. More theory can be found here.

https://people.duke.edu/~rnau/411searw.htm

Let me know if this helps or i can provide some detail.

#3

Hi @waparna, I’m also having the same challenge with a time series data? I’ve a 37 month data and its’s having a sudden peak in 35th month. I’ve taken this in my test data which results in poor accuracy. Arima(0,0,0)(0,1,0) appears to be the best model according to auto.arima?
Could you please brief as to what it means and how you approached this problem?

Many thanks!

#4

HI,

You can first take a log transform of your series to penalize the high peak in your data. Post that, apply autoarima to check which combination of p, q, d forms the best model.

Also, while applying auto arima, you can tune the parameters, such as set the seasonality, choose a stepwise model, set min max values of P,Q,D as well.

#5

Hi,
I have transformed my series using log and diff to make it stationary and now I want to forecast using the model but the otuput is coming as transformed i.e. original series was like 10000 and forecast is coming like 0.1 How to transform back into the original series so I can get the forecasted values in thousands.
Spareseries is my timeseries whcih consist numbers ranging between 10k and 15k
Code :
Spareseries2<-diff(log(Spareseries),12)
plot(Spareseries)
plot(Spareseries2)
Spare2 <- auto.arima(Spareseries2,trace=TRUE,test=“kpss”,ic=“aic”)
summary(Spare2)
confint(Spare2)
Spare_forecast2<-forecast(Spare2,h=3)

#6

Take exponential of the resulting data

#7

Ok that would suffice for logarithm but diff should not be treated separately or code do it internally