Statistical Demand Forecasting: Some Things To Consider
While statistical demand forecasting provides a more sophisticated method of predicting future demand, it does have several prerequisites in order to produce good results. First, it is important to understand what’s considered “good results”. When forecasting any demand stream, the more summarized the data, the greater the accuracy. For example, yearly item category level history will be far more accurately forecasted than monthly item by customer demand. This is simply because the information has most of the low-level noise averaged out of the demand stream.
During a recent APICS presentation, it was stated that best-in-class forecast accuracy at the product family or product category level is 90%; and at the product mix or item level it is 85%. These best-in-class values are difficult to attain across all product lines depending upon the amount of data and broadness of use (sparsity) of the product mix. The deeper in the product hierarchy you go, the less accuracy you can expect.
Important Characteristics of Historical Data in Demand Forecasting
Obviously statistical forecasting depends completely upon historical data with three important characteristics – good quality, the necessary amount of history, and the type of data used as inputs to the statistical engine. Besides the possible inaccuracies found in some data, poor quality is sometimes represented by historical changes in business conditions that produce data inputs that are out of context or inconsistent with future trends. An example could include a product line that was recently acquired.
In terms of the amount of historical data used, two years is typically the minimum in order to forecast trends and seasonality. Even with two years of historical data patterns and trends, the true data can be disguised in noise – resulting in inaccurate forecasts.
The type of data being forecasted is very important since the statistical engine will attempt to create a forecast with the continuous stream of year-over-year data supplied to it by comparing similar time periods from one year to another. Many times source input data has a mixture of information buried inside it such as customer order demand and supply chain performance. This is especially true in a supply-constrained environment that can produce inaccurate and misleading results relative to what the future demand plan should be built upon.
The Best Solution for Accurate Demand Forecasting
Even with all these constraints or possible hurdles that need to be overcome, statistical forecasting is the best solution to provide a baseline forecast for supply chain planning. It removes the bias, human error and provides a repeatable, automated method of demand forecasting for many product combinations. Imagine a statistical forecasting system as the “house” in a gambling casino – in the end, you won’t beat it!
With all this in mind, what then is the best possible solution for accurate demand forecasting? The answer is known as combination or collaborative forecasting. Such a solution combines baseline statistical forecasting results with overrides from users who have factual inputs about future events that are not represented in the historical data that is used to produce the baseline forecast. The process must be carefully managed and the appropriate users must be chosen carefully, but studies have shown this approach to be the most effective demand planning solution.
The Silvon Stratum Forecasting application will support any and all of these methods; but it still remains at the mercy of input data quality.