# MAPE is high for daily sale prediction

I have daily sales data from 2011 to 2013. I have to do prediction for 2014.I have used arima and exponential method to predict the daily sale, but it is not giving the better result. MAPE is around 25%.

y=ts(x,frequency=7)

fit <- auto.arima(y)

fc <- forecast(fit, h=265)

plot(fc)

fit <- ets(y)

fc <- forecast(fit,h=265)

Below is the link for data:

Is there a way to improve the MAPE? I am new to time series,I would appreciate any kind of help.

• What country is the data from as holidays can have an impact? What type of data is it? – Tom Reilly Jul 20 '14 at 13:33
• It is a daily sales data of a store. The country is US. I tried using holiday also. Still getting the same MAPE – Arushi Jul 20 '14 at 16:44

I looked closely at your data and the 2014 data is much much lower and couldn't have been predicted to be that low. Every January to March is high in the historical data and the model can't be blamed for this massive shift downwards. What you need to do is consider the MAPE at many different origins not one. When you do that you will see that as the first observations in 2014 data are lower that the model respond to the lower level of the data and your MAPE will get better.

• "MAPE at many different origins not one", could you please explain this sentence? I am getting it. Thanks. – Arushi Jul 22 '14 at 8:02
• It means that you need to look at forecasts from today, tomorrow, the next day and then calculate a WMAPE(weighted) and not just from one point in time. – Tom Reilly Jul 23 '14 at 13:41