I have monthly sales data for about 7 years and i need to forecast sales figures for the next 10-15 years (yearly basis, not monthly). Also i need to use ARIMA for this.
How to approach this task properly? Should i build a model for the monthly sales figures, forecast all months of needed 15 years and than aggregate it to the level of years? Or is it better to convert my monthly time series to yearly and then make a forecast (however it will be only 7 data points for the needed forecast of 15)? Or maybe some combination?
Best Answer
"All models are wrong, some are useful." George Box said this and he created ARIMA models. ARIMA will work just fine. ARIMA models are a superset of all other models(exponential smoothing, etc.) except H-W Multiplicative. If you have a causal variable you will enhance your model and forecast. If you think the sales are tried to population or some other variable then by all means use it. If there is a life cycle to your product where it will become obsolete in 3 years then you can guide the forecast down by creating a dummy variable with a 1 in the history and right before the drop and then using .9, .8, .7 etc for example to guide the forecast down. You can also use analogues as causals to do something similar.
Don't sum your data to annual. Just sum the monthly forecasts to yearly when you are done.