Weekly sales forecasting using time series decomposition with irregular seasonality: A case study of Sony TV sales in the United States

In the practice of sales forecasting, alongside high accuracy, simplicity and interpretability are desirable qualities in a predictive model. Here, we use time series decomposition to construct a simple and interpretable forecasting method for weekly sales, taking as a case study the total weekly sales of Sony TV units in the United States. The main challenge is to properly address irregular seasonality arising from the way a year is divided into weeks and from floating holidays and events. As a solution, we propose the concept of ‘analogous weeks’, a general framework for grouping weeks that are similar according to some criteria, which encodes seasonality and already accounts for calendar effects. By incorporating this notion into the time series decomposition and forecasting, we achieved yearly average percentage errors of less than 10% for the 1-week-ahead forecast and less than 15% for forecast horizons of up to 26 weeks, accuracy levels comparable to or better than those of other established seasonal models, namely Prophet and SARIMAX, particularly at longer forecast horizons.