# What time series model should be used?

Given the following daily time series data I have used auto.arima in R to build a model. I used freq = 5 because the data is collected only on weekdays (Monday through Friday).

tsdata <- ts(data, freq = 5)
fit <- auto.arima(tsdata, ic = 'aicc')
plot(forecast(fit, h = 80))


return,

Series: tsdata
ARIMA(3,1,2)(0,0,1) with drift

Coefficients:
ar1      ar2      ar3     ma1      ma2     sma1    drift
1.6825  -0.6384  -0.0993  0.0151  -0.9778  -0.0063  -0.0232
s.e.  0.0494   0.0904   0.0487  0.0125   0.0123   0.0483   0.2080

sigma^2 estimated as 33.92:  log likelihood=-1393.24
AIC=2806.1   AICc=2806.44   BIC=2838.78


I obtained the following forecast result: Why does Arima only give me back a straight line ?

Is the ARIMA model is the correct one to use here ?

How do I retain the feature of the forecast data ?

I am very new to build time series model.

Any suggestions would be appreciated.

• You could also try the ets function and compare the results. Also try fit <- auto.arima(tsdata, ic = 'aicc', stepwise=FALSE) – Zach Apr 2 '14 at 14:57
• auto.arima is not suitable for high frequency data, try stlf function in the same forecast package. – forecaster Apr 2 '14 at 16:11
• The data are daily with frequency 5. That isn't high frequency. – Rob Hyndman Apr 2 '14 at 20:49

## 1 Answer

There's quite a lot to fitting an ARIMA model properly. As with most statistical models there are things you need to do to check that the modelling assumptions are correct and that you've chosen the most appropriate form. The auto.arima makes some sensible choices for you but to understand these choices you'll need to do some background reading. Here's a useful place to start http://people.duke.edu/~rnau/411arim.htm

However a short answer to your question is that there apparently isn't much by way of trend or seasonality in your time series and therefore the best forecast is a straight line. (It's often difficult to convince people that straight line is the best forecast - they want to see something complicated looking in order to be convinced).

• What about I use neural network on this data ? any neural network package I can use in forecast ? Thank you – newforecaster Apr 3 '14 at 4:05