Companies use forecasting for many purposes, ranging from financial budgeting to production or labor planning to inventory management. Each of those decisions may require a different type of forecasting. So before jumping into the data and model, it’s important to understand first the purpose of the forecast and how it can improve decisions. To do that, it’s useful to ask the following set of questions.

What is the role of uncertainty?

Uncertainty pervades all forecasts, but the level and importance of uncertainty will vary among different applications. For a market growth forecast, uncertainty informs the risks for each strategic choice, so we need to incorporate it thoroughly into the forecasting process. For inventory management, uncertainty of demand bears on whether a company will meet the desired service level. For pricing, decisions usually come at an aggregated level, so the size of uncertainty shrinks and is less important. Forecast results could consist of a point forecast with no uncertainty, probabilistic results with several prediction intervals or scenario-based forecast results corresponding to different levels of uncertainty.

What demand are you trying to forecast?

Although we commonly talk about demand forecast, most of the time we cannot observe the true demand because of real-world limitations. A retailer loses sales due to stock-outs or poor assortments. A manufacturer does not fill every order due to production capacity limits or scarce parts. Therefore, the quantities we observe, such as orders or shipments, might be less than the true demand.

Although true demand is hard to observe, sometimes we can estimate it by combining different observed quantities and then forecast such estimated true demand—that is, an unconstrained forecast. At other times, we do not need to forecast the true demand; we only need to forecast the observed constrained demand—that is, a constrained forecast. For instance, production planning merits an unconstrained forecast because the organization can take actions to meet the true demand. By contrast, a constrained forecast makes sense for sales targeting or financial projection because there is no point in planning for demand that you cannot fulfill.

We recently helped an industrial equipment company improve the planning process for producing and subleasing equipment. Previously, the company had forecast the probability of sales based on project history, which consists of realized opportunities—constrained demand. By shifting to a forecast of the probability of sales at different levels of opportunity, the company realized millions of dollars in savings through improved equipment utilization and freight and inventory management.

What information do you need to consider?

For a market growth forecast, external factors such as the macro economy, government policies and competitors’ decisions tend to have a big influence. For revenue management, internal factors such as pricing and promotions will combine with external factors such as competitor behavior to influence revenue. For operational level forecasts, the internal factors typically are most important.

How do you want to interpret the forecast?

For pricing, the effect of price on demand or sales obviously is critical for decisions, so the forecasting should include an explicit model to capture these relationships. For inventory, the factors that drive demand usually do not affect the decision, so a company could use either interpretable models or black box models.

What forecast values are preferable?

For intermittent demand—that is, low demand interspersed with periods of zero demand—the forecasts could be either a mix of zeros and nonzeros or stable small values close to zero. For merchandise planning, either way would work, but for inventory management, the latter usually works better since it would make replenishment easier and reduce costs associated with replenishment variability.

How do you measure forecast performance?

Forecast performance measurement only makes sense at the level of the decision. This applies to the hierarchical level—say, inventory replenishment at the SKU-in-store level. Due to the noisy nature of the data at this level, we might model it at a slightly higher level in the hierarchy, such as the SKU-in-region level, and then disaggregate the high-level forecast down to the low level. In this case, measuring the forecast performance at the modeling level has no actual meaning; instead, it is more of a by-product in the forecasting process.

Turning to the time frequency level, an example is labor planning for each shift. Sometimes we model the demand at a different frequency: If the shift is eight hours, we might model hourly demand, whereas if the shift is weekday vs. weekend, we might model daily demand. Then we convert the demand forecast information at the modeling time frequency to the shift time frequency, using peak demand or average demand methods during the shift period. Forecast accuracy during the modeling time frequency is less useful than accuracy during the time frequency regarding which executives will make labor-planning decisions.

The central point to keep in mind is that forecasting can only help improve business decisions if the forecasting method suits the particular decision at hand.

Yue Li is an expert with Bain & Company’s Advanced Analytics practice. She is based in Los Angeles.

Read More

Advanced Analytics Expert Commentary

Success with advanced analytics requires both technical know-how and a thoughtful approach. In this series, Bain's experts offer practical advice on some of the most common data issues.