The Statistical Forecasting Conundrum

No other component gets more attention and, dare I say, abuse than the forecast itself. When we run out of stock or miss service levels, it’s the forecast that takes the blame. When we are overstocked, it’s because the forecast is too high. Even when we tweak it, things never seem to get better.  The reason is simple. Today, most traditional systems only focus on one segment of forecasting, while the technologically advanced systems have the ability to generate a holistic view of the forecast.

What are those different segments of the forecast? In previous articles we have defined stochastic, deterministic and causal forecasts, but today we will concentrate on just one: the stochastic forecast. Additionally, we will discuss the inherent pitfalls of managing this portion of the forecast.

The forecast is more than just the weekly, or 4-weekly number that you see on your demand-forecasting screen. In fact, that is only a small portion of the entire forecasting picture. The forecast in Blue Ridge’s Supply Chain Planning system consists of 3 different segments:

1. Stochastic Forecast – also known as the statistical forecast
2. Causal Forecast – this is the segment of the forecast that we can identify where the demand came from and/or why it happened (i.e.: seasonality, promotions, weather, etc.)
3. Deterministic Forecast – this could be from orders that we’ve pre-booked or previously committed to, scheduled orders or even demand that is coming from the lower echelon in a multi-echelon environment

So, what really is the stochastic forecast? What does it mean? By definition, the word stochastic means randomly determined or a pattern that can be analyzed statistically, but may not be predicted precisely.  Wikipedia describes stochastic forecasting as an approach that involves both formal and informal methods- using educated guesses and quantitative methods like sales data and statistical techniques.

To put in laymen’s terms, we really don’t know why the demand occurs. Since we don’t know why it occurs, the stochastic forecast relies on historical data to predict the future. It assumes that what happened in the past will repeat itself in the future.

Most traditional systems rely on this segment to generate the forecast. The problem with only considering this segment is that it gives no consideration to demand drivers impacting the item/SKU. So, when history does not repeat itself, the result is a highly inaccurate forecast, missed service, lost sales and angry customers.

This problem is magnified when the human element is inserted, and intuition, opinion and “market knowledge” take over. The forecaster starts off fighting the urge to react emotionally, but eventually emotions take over and lead to an overreaction. This results in the very same things we are trying to avoid: out of stocks, poor service level, overstocks and huge swings in cash flows.

The hard truth is that stochastic/statistical models were created long ago when today’s computing power didn’t exist. Even more recently, when forecasting methods first became automated, the computing power, memory and data storage costs were too high. Using old-school replenishment systems, planners were forced to accept the presumptions made in the random forecast and rely on the outdated methods. However, because of the processing power available to us today, we can now use the wealth of data to segment the demand and achieve a more precise level of forecasting.

The best way to approach the forecast is to analyze each item individually and segment demand based on causal factors, deterministic considerations, and finally, on stochastic demand for every item in each location. While unrealistic and unattainable in the past, this is the reality of today’s technology. By capturing customer transactions and the influencing factors, it is possible to gain both a stochastic and causal understanding of demand. To the extent that causal factors cannot be identified, or don’t show substantial influence on demand, that portion of data is relegated to a statistical forecast. Demand is segmented to apply the most effective technique, ensuring that the combined forecast is as accurate as possible. In-depth demand segmentation reduces the variability of demand and provides insight into demand-driving factors.

Relying solely on a stochastic model for a demand forecast is a mistake. In addition to all of the challenges in today’s marketplace, the risk associated with a purely statistical forecast is multiplying your problems. At Blue Ridge, we include all three segments of demand and give our customers a comprehensive view of the forecast. Reach out to us for more information on the pitfalls of statistical forecasting and we would be happy to discuss.