Part III: Important considerations in Forecasting
This is the last blog on Demand Forecasting series and we use this to write on the topics we left uncovered (seasonal and intermittent items) and also to enumerate some of the lessons we have learnt overtime.
Complexity versus flexibility: In evaluating a forecasting solution, we recommend a retailer to thoroughly investigate between complexity and flexibility of the software. We define complexity as the advanced nature of analytics in the solution and flexibility as the degree to which the solution can be customized. We have found the retail packages of many software providers offering more complexity at the cost of flexibility; whereas we rate flexibility over complexity. If one believes that every retailer, store or product is unique, one can easily come to agreement with our conclusion. Advanced analytics has merits but with high degree of customization, accurate forecasts can be produced even with relative simple statistics. Such solutions, moreover, are easy to communicate from business context and have important side-uses (Check out our blog Promotion Planning and Analysis)
Art versus science: Accurate forecasts combine intelligent algorithms with feedback from merchandisers and store-managers. Achieving this balance, however, is easier-said-than-done. The power of information outside the database in improving forecast accuracy should not be undermined. With a highly customizable forecasting system such information should also be easy to incorporate. Demand for items, for example, rapidly drops beyond a certain inventory on display, which in-turn is dependent of the size of the display. In other words, customers refrain from buying products when they see relatively empty shelves. Input from a store manager that the display size for a particular store is smaller or larger than other stores could be an invaluable input to improve the forecast accuracy at that store.
Timeline for historical data: We have come across arguments between using two and three years of historical data. The trade-off here is between robustness of forecasts (three years data) and recency (two years data). At least 2 years of historical data are required for robust seasonality calculations and is generally sufficient in many cases.
Improvements to cannibalization: We propose an improvement to the cannibalization modeling as mentioned in our previous blog Demand Forecasting Part II: How do I model cannibalization between similar skus, items or products? Although Merchandise or Store Managers are better than algorithms in identifying close substitutes, best is input from customers. We take a cue from Amazon’s People who viewed this…approach and recommend using customer browsing data from the firm’s online store as it provides a direct feedback from users.
Intermittent and seasonal items: With low sales and activity (sale frequency) forecast accuracy is severely hampered, especially at the store-level where possibly a handful units of an item are sold over several weeks. The aggregation of large forecast errors across stores, may lead to wildly inaccurate forecasts at the chain-level and hence inaccurate replenishment or order decisions. The retailer needs to carefully weigh the benefits of a store-level forecast in such cases and where possible should instead focus on regional- or distribution-center-level forecasts. An allocation to divide the regional inventory between stores could be made based on Store or Merchandise manager’s inputs. Moreover, inventory at distribution centers can easily be shipped to stores with a very low lead time.
Large volume seasonal items, for example: marshmallows, can be forecasted using parameterization of the historical sales curve around the seasonal peaks.
By: Sagar Kalyankar