1
$\begingroup$

Ok, so I understand that it's impossible to capture seasonality or trend from only 1 period since there is no baseline to compare to.

I believe a great example I read was: "try to guess the next number in the sequence, 2,...."

However, what about when you have a partial set of data for the next period? E.g. You are trying to forecast monthly sales data and have 16 months worth of historical data. Let's say from Nov'15 to Feb'17.

Why can't you capture the trend from let's say Feb'16 and Feb'17 and use that as a baseline assumption in combintation with the March'16 value for a March'17 forecast?

What other approaches can I use to help solve this forecasting scenario (it seems to me all seasonal models require a minimum of 2 periods to work)? In another thread, someone referenced using a TBATS model however that just gives me a straight line for a forecast which is not too desirable. The only other thing that comes to mind is a seasonal naive.

$\endgroup$
1
$\begingroup$

A standard approach here would be to include the sales from the same period in the previous year as a parameter for predicting sales for a period in the future. So instead of a parameter for each month separately, you just have a previous-year-sales parameter. You would of course also include other relevant quantities in the model, such as marketing spend.

This will automatically handle dips or increases in sales over holiday periods, as long as they align with the months reasonably well. If holidays greatly affect your sales, you would want to model which month they occur in each year. This is more important with a daily model than a monthly model though.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.