How Much Data is Needed to Evaluate Trade Promotions?



To answer this question, we need to start with the type of model needed.

 If you are using an exponential smoothing model very little historical data is needed.  One of the advantages of an exponential smoothing model is the lack of a need for significant historical information. If you use this type of model the amount of historical data can be from 1 to 24 periods and more likely weighted to shorter rather than longer.  Exponential smoothing models also don’t require the maintenance of an historical database since it maintains a single number that represents the weighted average of the historical numbers.

It is more difficult to determine the length of the historical database for a univariate model.  Since the variables are considered one at a time, how they are backed out is as influential as the length of the database.  There are no good scientific or mathematical studies addressing this issue as it relates to Trade Promotions but anything less than 2 years is not going to give good estimates so the normal estimate of what is needed is 2 to 4 years.

Five years are recommended for the multivariate models like the SPARLINE Model.   Five years is recommended to allow sufficient history to calculate INDUCED SEASONALITY and to be able to factor in the impact of other marketing variables that have an impact on sales.

Five years is recommended because that is the optimal length balancing the amount of historical data with the need to have enough stability in the data to accurately calculate a SPARLINE or baseline and by generating such variables as seasonality and trend.  If you are tracking batteries and have a hurricane one year it will be very hard to get seasonality because the one year is so distorted.  And if something happened in the second year artificially increasing sales, seasonality would be significantly distorted which would lead to inaccurate analytics.  A two-year study was conducted in which many 100’s of products forecasted future sales using 3 to 10 years of historical data.  Little change in the accuracy occurred by using 5,6,7,8,9, or 10 years of history.  With four years of history the accuracy started to drop and with 3 years there was a significant reduction in accuracy.  Therefore 5 years is recommended.  However in the real world you can’t always get 5 years, and for new products sometimes only a few months of sales exist.  Since you have to forecast all products, the SPARLINE MODEL has special algorithms to handle short history items where it is impossible to calculate factors such as seasonality.

Time periods.  It is recommended data be saved weekly.  Some companies save their sales data by the company’s fiscal periods but this is not very helpful.  Promotions are run based on weeks.  In order to do a sound analysis, the promotions have to be evaluated based on the sales covering the promotion period.  Daily data can be saved but this has little incremental value and unless the Company is saving daily data for other reasons the extra storage and processing don’t justify doing that for trade promotion analysis or forecasting.  Many other market factors also act weekly (e.g. the short-term impact of price increases) and not quantifying the impact by week will lead to material errors in the analysis.