Forecasting 101

The primary building block of any integrated business planning process is a reliable demand forecast. A statistical forecast is generated through detailed analysis of historical time series data. If the demand data has been compromised by anomalies, older statistical models may project and even amplify the effect of erroneous data points into the forecast. Intervention - whether manual or automated - is prone to error, resulting in increased costs and disappointed customers.

Definitions

  • Time series: A sequence of observations measured at uniform time intervals.

  • Forecasting: The process of making statements about events whose actual outcomes have not yet been observed.

  • Time series forecasting: The use of a model to predict future values based on previously observed values.

Basic Elements

  • Level: Usually a weighted moving average of data points projecting a steady-state (flat line).

  • Trend: Projects the rate at which the data is increasing or decreasing (sloped line).

  • Seasonality: Projects the magnitude and nature of patterns repeating in a uniform cycle (undulating line).

Challenges

A good demand forecasting system should minimize inventory and production costs while maximizing on-time-delivery to the customer. The objective is to find a cost effective and user friendly forecasting system that minimizes forecast error.

Traditional models such as Holt-Winters and Box-Jenkins have been staples in statistical forecasting systems for many years, but are somewhat rigid in methodology. Holt-Winters uses a technique called exponential smoothing to gradually incorporate new data into the projection. The smoothing coefficients are often set arbitrarily and tend to allocate disproportionate weight to the initial data point. Box-Jenkins applies an autoregressive moving average to try and most accurately reproduce the historical data. This backward-looking technique does not necessarily identify factors that telegraph changes in the projection looking forward.

Common Errors

Sporadic, anomalous or limited demand history can cause false estimates of the level, trend and seasonality, resulting in an inaccurate statistical forecast.

A forecast that is too high (over-forecasting) can result in excess spending on inventory and production. A forecast that is too low (under-forecasting) can result in poor on-time-delivery and customer satisfaction. Errors generated by forecast software due to anomalies are often addressed in the following ways:

  • Manual Selection: User attempts to select an algorithm that minimizes exaggeration of anomalies. This typically involves visually analyzing the forecast produced by each algorithm individually and identifying the one that appears to best extrapolate the demand history. This process is extremely time-consuming and the user cannot be certain that their algorithm selection is appropriate.

  • Manual Adjustment: User adjusts demand history to try and eliminate the anomalies causing errors. This means that the user attempts to correct the demand history by identifying and changing the values of data points that do not appear to align with the rest of the demand history. This is a challenging and very subjective approach and is thus prone to inaccuracy.

  • Automated Adjustment: System set to automatically adjusts “outliers” identified in the demand history. This approach relies heavily on assumptions regarding what constitutes an outlier, as well as how best to correct it. Once again, it cannot be guaranteed that the outliers identified are true anomalies, or that their adjustment is appropriate and will result in a more accurate forecast.

If these techniques are unsuccessful, the user will be required to manually override the statistical forecast, thus negating its purpose.