Part I: How much do I order given point-of-sales, promotion history and seasonality patterns?

Demand forecasting is as much an art as a science. By this we mean that close collaboration between assortment planners or merchandisers and the analytical team is paramount. We will cover this aspect in greater detail in Part III: Important considerations in Forecasting. Please browse our case studies for live examples.

Demand forecasting involves three major steps – identifying skus or items which would require special handling, developing causal forecast for remaining products and fine-tuning the model.

Demand for items with very low and/or intermittent sales is very difficult to forecast through causal models. Such items can be identified by parameterizing the historical sales curves for such items (a rather complex process) or using some combination of average sales and number of weeks with positive sales (straightforward process). Advanced pattern recognition processes could also be used for this purpose. Refer chart below for an illustration. Items with very low sales or frequency require special consideration and are discussed in the Part III: Important considerations in Forecasting. We call the items with high enough sales and sales activity as Regular items.  

Demand for Regular items is forecasted using causal models. Literature suggests using causal models in a top-down or bottom-up hierarchical forecasting process. Top-down forecasts involve forecasting at an aggregate level – for example: at the chain-level for an item. An allocation algorithm is used to spread the forecast at the lower levels – for example: store or day level. Contrastingly, bottom-up approach starts from the lowest level – for example: sku-store-day and then aggregates the forecast to higher-levels for decision-making. Based on our experience with various retailers, however, we recommend a hybrid of the two approaches. This approach involves forecasting at the lowest possible level where possible and aggregating at the next higher level if necessary. For example, if the forecasts are not significant at the sku-store-week level, we will develop the forecast at the sku-region-week level. To forecast at the lower level we then either use the causal parameters (regression betas) at the higher level or pro-rate the forecast at the higher level at lower levels. The level at which we forecast depends on the statistical models (science) as well as subjective business knowledge (art).

Broadly, causal models involve two components – seasonality and promotional elements. Seasonality can be estimated using dummy variables in the regression for days or weeks (simpler process) or using spectral density methods. Promotional elements include discounts and promotions like weekly or seasonal catalogs, display in stores, and media like television advertisements or events like Family and friend’s sale and cannibalization. Cannibalization is the effect of promotion on one item on the sales of similar item, for example: Coke and Pepsi cans. Cannibalization is an advanced topic and is dealt separately in Part II: How do I model cannibalization between similar skus, items or products? The regression model is used at successive levels in the product-calendar-store hierarchy till a statistical significant and logical model from a business standpoint is obtained.

The last step - fine-tuning involves comparing the forecast produced above with the testing or holdout sales, and re-adjusting the average sales and sales activity parameters in step one and remodeling the regression inputs in step two as needed. Trend component may need to be identified depending on the demand pattern. Regular items with weekly sales greater than five at the store-level should have forecast error around 30%, as a rule of thumb. This been said, the errors could vary significantly based with specific cases.

By: Sagar Kalyankar


Add new comment

Filtered HTML

  • Web page addresses and e-mail addresses turn into links automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <blockquote> <code> <ul> <ol> <li> <dl> <dt> <dd>
  • Lines and paragraphs break automatically.

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.