# Constant forecasts in SPSS

I have weekly data for the last four years. I am using SPSS to do forecasting. I am getting a constant value in the forecast period. What could be the reason behind it? Is it due to defining weekly dates? I am not able to figure it out.

I have defined the week dates using following code:

DATE WEEK 1 52

• Your fit models the spikes very well. I assume you fed it some kind of explanatory variable, e.g., a Boolean regressor coding a promotion. If these explanatory variables are constant in the forecast period (and your model does not include trend, seasonality or ARMA terms), then the forecast will naturally be constant. Does this help? If not, it would be good if you could edit your question to explain what forecasting model you are using. – Stephan Kolassa Mar 27 '14 at 10:48
• No I haven't used any explanatory variables. I have juts given the input for independent variable. Secondly, I am using Expert Modeler option in SPSS which has given me the ARIMA(0,0,1)(0,0,0) as the best model.. – Nevedita Mar 27 '14 at 11:05

SPSS sees an MA(1) process (that is the (0,0,1) term) and no seasonality in the data (that is the (0,0,0) term). An MA(1) term has very little impact on the point forecast. Here is an example in R with toy data, where I artificially inflate every 13th observation. Actuals in black, in-sample fits in red. Note how the forecast in blue is almost flat, although the series and the fits show clear spikes:

R code:

library(forecast)
set.seed(1)
xx <- rnorm(200)
xx[13*(1:15)] <- xx[26*(1:5)]+8
model <- arima(x=xx,order=c(0,0,1),seasonal=list(order=c(0,0,0),period=52))
model

plot(forecast(model,h=53),ylim=range(xx+model$residuals)) lines(xx+model$residuals,col="red")


ARIMA models can be extremely unintuitive. I suggest you play around with simulated ARIMA(p,d,q) models to get a feeling for them. I always recommend this textbook to learn more about forecasting.

To return to your data: you have giant spikes there, and an ARIMA model won't be able to explain those and forecast them. I suggest that you go back to the data source and see whether you can find some explanatory variables (e.g., promotions) to build a causal model. Such a model would allow you to forecast the spikes in the future again.