Forecast FAQ

Why is a Forecast important to VMI?

Forecasting provides an estimate of what demand levels might be expected based on historical patterns.

Forecasting also provides a statistically valid measurement of the range of error likely with the forecast.

Together, this data can be used to drive various economic evaluations and decisions.

Forecasting can also help to identify demand patterns where forecast is not usable.

What Type Of Demand Patterns Must Be Anticipated?

Demand patterns vary based on the industry and channel. Within the Commercial / Industrial channel, which is our focus, demand patterns vary greatly from item to item, but show many similarities across industries / channels.

• High volume to low volume
• Frequent demand to sporadic demand
• Many customers to few customers

• Many items must be managed, so automation and scalability are critical
• VMI is dealing with independent demand.
• Many, and in some cases most items will have demand patterns that will not create a reliable forecast.
• For some items, the replenished location will choose to not stock, but rather to order only to fill backorders.

What Forecasting Models Are Applicable?

• Used on very short data series where it is not feasible to fit more complex models to the data
• These forecast consist of the un-trended, non-seasonal n-term moving average, often with multiple possible values for the n-term.

• Used for demand series with numerous zeros
• This model produces a straight horizontal line forecast, which is often superior to those from exponential smoothing for low volume, messy data.

• Discrete (Negative binomial and Poisson distributions) are typically used on data whose values are small integers. The model can consider and forecast both trend and seasonality.

• A 'family' of methods which requires longer data series, they are typically the most widely applicable of the time series methods to business data.
• A good forecast process might consider nine models of Holt-Winters exponential smoothing, i.e. the combinations of Constant Level / Linear Trend / Damped Trend with Non-Seasonal, Additive Seasonal and Multiplicative Seasonal. In an advanced system, the alpha values for level, trend, seasonal and event, as well as the 'decay constant', are variable, specifically fitted to the data.

With this range of possibilities, a method is needed to allow automation of the model selection. We select the 'winner' of the nine models based on Bayesian Information Criterion (BIC), which is a goodness-of-fit criterion.

Why Is Forecast Automation Important?

With VMI programs managing thousands (or 10's / 100's of thousands) of item / location combinations (or more), automated forecasting is imperative. However, automation must not be allowed to create unacceptable replenishment risks. Thus, VMI programs often surround the forecast process with advanced demand assessment features on the front end and with replenishment model selection features on the back end.

What If My Transactional Data Contains Errors?

Perhaps most importantly, data is not always 100% correct. Even with methodical testing up front, errors can occur in a VMI process. Advanced VMI systems recognize this fact and contain fail-safe techniques to limit the exposure that can be created when incorrect data enters a forecasting process.

Is There A Formal Process For Evaluating Forecasting Systems?

Yes. The M-3 forecasting competition was designed to evaluate the accuracy of different forecasting methods. The competition was sponsored by the International Journal of Forecasting and is the largest and most comprehensive empirical forecasting study ever performed. Professors Spiros Makridakis and Michele Hibon of the French business school INSEAD reported the results.

The forecasting engine (now licensed by Pan-Pro and integrated into the Pan-Pro VMI system) significantly outperformed all of the other software approaches as well as 18 out of 19 academic teams.

The M-3 competition is the second time that the Forecast engine has outperformed human practitioners and other forecasting packages in a formal study. The first time was in a much smaller study conducted by Keith Ord and Sam Lowe of Penn State and published in the February, 1996 issue of The American Statistician.

This forecasting process is suitable for other, non VMI forecasting as well.