Demand Forecasting in Retail

by Michael Gilliland

Modern demand-forecasting systems provide new opportunities to improve retail performance. Although the art of the individual merchant may never be replaced, it can be augmented by an efficient, objective and scientific approach to forecasting demand. Large-scale systems are now capable of handling the mass of retail transaction data – organizing it, mining it and projecting it into future customer behavior. This new approach to demand forecasting in retail will contribute to the accuracy of future plans, the satisfaction of future customers and the overall efficiency and profitability of retail operations.

Retailers face several challenges when it comes to forecasting:

• Scale of the problem (large number of stores and items to forecast).
• Intermittent demand (slow and erratic sales for many items at the store level).
• Assortment instability (frequent new-item introductions and seasonal assortment changes).
• Pricing and promotional activity.

Given these challenges, it is important to recognize where forecasting can enable better retail processes, and where forecasting alone will not solve the business problem.

Large-scale automated forecasting
The objective in any business is to have the right product in the right place at the right time – and in the appropriate quantity. Achieving this objective can be a significant challenge for retailers because of the sheer number of items they carry and the number of stores where the items are stocked. A large retailer may have tens of millions of store/item combinations.

Given this situation, it is clearly impractical to attempt to manually forecast demand for each item at each store. It would not be economically feasible to employ the hundreds (or thousands) of demand analysts necessary to manage each forecast individually. Fortunately, it is neither necessary nor advisable to manually create or intervene in each forecast at the store/item level. Large-scale automated forecasting software (such as SAS^® Forecast Server) can address this problem. In most situations a quality forecast can be created with little or no human involvement. This automation minimizes staffing requirements, while permitting forecasters to focus on the "high value" forecasts that have the greatest impact on customer satisfaction and financial performance.

Forecasting and revenue optimization
Revenue optimization systems help the retail planner make better decisions on regular product pricing, promotional activity and markdown pricing. Such systems are designed to optimize an objective (e.g., maximize revenue, maximize margin or minimize inventory). They succeed to the extent that they drive actions that improve financial results.

Optimization decisions are based, to a large extent, on the forecasted impact of various possible scenarios. For example, to determine whether it is better to price an item at $1.99 or $1.79, these systems must estimate the sales of product at these price points. These "forecasts" might be based on prior sales history for this item at various price points, or based on the pooled history of other items that have incurred price changes.

Forecasting and replenishment
Retailers want to grow sales, realize profits and satisfy customers. This means having stores (and warehouses) sufficiently stocked to meet customer demand, but not so overstocked that excessive capital is tied up in inventory or that significant season-end markdowns may be necessary.

Store-level stockouts have many possible causes. A poor forecast of demand (resulting in the item selling out) is one possibility, but there are others:

• Poor replenishment policy (failure to account for demand variability, supply variability, forecast error, etc. in making inventory plans).
• Poor replenishment practice (failure to properly execute inventory plans).
• Shrinkage (loss of sellable inventory due to theft, damage or misplacement).

A good replenishment policy takes into account the uncertainties of supply and demand, and makes store-level inventory less dependent on a highly accurate forecast. Accurate forecasting at the store/item level is inherently difficult due to the amount of volatility and randomness in demand at this level of granularity. Sporadic or intermittent demand can also be a major problem, as illustrated by one "top 40" retailer that reported sales of less than one unit per week for half of its 30 million store/item combinations. Pooling demand across stores and generating forecasts at a region or warehouse level can help solve this problem. Forecasts will be more accurate at the aggregated level, and attention can be focused on maintaining the appropriate level of inventory at the warehouse. Good replenishment policy and execution will allow stores to maintain appropriate stock levels, without overdependence on store- or item-specific forecasts.

Forecastability of retail demand
Forecasts will never be perfect, and sometimes they may not even be very good. The goal of forecasting in retail should not be a foolish pursuit of perfection, but to generate forecasts that are as accurate and unbiased as we can reasonably expect them to be, and to do this as efficiently as possible. Large-scale automation helps solve the problem of generating forecasts at granular levels of detail (such as store, item or week). However, there must still be a realistic assessment of the likely accuracy of forecasts at that level, and consideration of other strategies that can be used in conjunction with forecasting to best solve the business problem.