Knowing When (and When Not) to Use Statistical Forecasting
In order to strive for consistency, or because of system limitations, an organization will often use the same forecast method for every item. Proper analysis by the demand planner will often show a clear separation between items that should be forecast statistically and those that require manual input.
Even for statistical forecasts, a knowledgeable demand planner with the proper tools will know which parameters to tweak in the various statistical methods to give greater weight to recent history. This becomes especially crucial in the current market conditions.
There are many reasons why products and/or specific product/ship-to combinations might require manual input. A skilled demand planner needs to be able to recognize and deal with them. The following are a few possible scenarios:
- Isolate causes of variability: The adage "one bad apple can spoil the barrel" is often true for forecasts. For example, a product-level forecast may appear to be highly variable, but breaking out the data at the ship-to customer level may indicate that most customers for the product are very predictable. One or two customers with a large demand that varies widely may be the entire cause of the product variability. The ability to perform this analysis and to forecast the "problem" customers separately allows the demand planner to create a much better product-level forecast.
- Eliminate sources of bias: Bias can be defined as a tendency to forecast consistently too high or consistently too low. While it is possible to see bias as a result of personality, often it is organizational metrics, reward systems and pressures that create bias. For example, rewarding Sales for exceeding forecast will introduce a bias for forecasting too low. A bias toward over forecasting in the later months of the year is introduced if "catching up to the budget" is allowed to influence the forecast.
- Collaborate with multiple sources: A forecast is only as accurate as the data that go into it. Often, the data required for developing a forecast are in multiple places. Sales may have some information; Customer Service may have additional pieces, and so forth. A robust system to collect and reconcile the different inputs is critical to arriving at an accurate forecast.
A good demand planner understands that all these inputs are important. She works with different areas of the organization and makes data available to them at the appropriate level. For example, Sales may want to look at it by product/ship-to; Marketing might want it by product family. Based on forecast accuracy metrics, the demand planner defines a way of combining these inputs into a final number and gaining consensus for the final forecast.
Often, even when everyone is using the best estimate, the sum of the parts is too big for the whole. If each of 20 sales reps forecasts just a bit too optimistically at the product/ship-to level, the result at the product level is a number that is not even in the ballpark. The demand planner should be able to recognize this (whether on the high or low side) and go to the appropriate people for a reality check.
Picking the Right Metrics for the Right Audience
Too often, an organization tries to reduce its forecast accuracy to a single number or even one number per product family. This is meaningless, because it allows the "too lows" to wash out the "too highs."
Naturally, high-level executives shouldn't have to look at the same detailed forecast accuracy metrics that a sales rep should look at. It is a key skill for the demand planner to recognize this and to present the right metrics for each audience. Executives, for example, need at least three metrics: one to show directional trend (getting better, getting worse), one to show bias (consistently too high or too low), and one to show magnitude. The latter metric means that a 40 percent forecast error on a $10,000 product line is not nearly as significant as a 20 percent error on a $1 million dollar product line.