A Quick Guide to Business Forecasting
Supply Chain Management often begins with a forecast. Two common types of forecasting methods are time-series and regression. In a time-series forecast, you are using past data points of the variable you are forecasting to predict the future of the variable. By contrast, in a regression, you may be using one or more other variables that you believe are correlated with the variable to predict the target variable. The exception to this is the ARIMA (auto-regressive integrated moving average) model which uses a variable’s own data history to regress upon itself. A good understanding of statistics is required to build a proper regression model.
Besides for using the previous period’s outcome as the next period’s forecast, the moving average model is the most basic of all time-series models. The moving average is using an average of a predetermined number of the variable’s previous outcomes to forecast the next period’s outcome. The number of periods used is dependent on the forecaster’s assessment of how far historical data is relevant for the future. A weighted moving average will typically weigh more recent data points heavier than data points further in the past. Exponential Smoothing is a variant of the weighted moving average where the weights used are mathematically expressed.
You may also incorporate seasonality into the forecast. For example, in a retail environment, demand usually peaks in the periods before the holiday season. These peaks and valleys often display consistent patterns year over year that can be mathematically traced and adapted to the forecast model to improve forecast accuracy.
A linear regression (also called ordinary least squares) requires both statistics and deeper business knowledge to produce. Regression creates a straight line that describes the relationship between the explanatory (independent) and response (dependent) variables by minimizing the sum of squared errors between the line and data points. You may use demographics as an explanatory variable for a certain product demand forecast. For example, you may use the number of infants in the area to predict the demand for baby formula and diapers for your store since they are used by this demographic. A multiple regression will use two or more explanatory variables to predict the response variable. For example, you might add economic data like GDP as another explanatory variable to the previous example.
This next point has been endlessly and often annoyingly brought up but it is such an important point that I have to mention it. Regression analysis only shows the correlation between variables and not causation. The independent variable may in fact cause the dependent variable, but the reverse can also be true. Additionally, they may or may not have a direct relationship at all. A third or multiple other variables outside the model may be causing the correlation. This fact has implications for demand management which I will mention in the next section.
No forecast is ever correct, they just estimate so that you are not randomly throwing darts at a board or using a Magic 8 Ball. Luckily all these mathematical models can provide error measurements between forecasted and actual data to give you a sense of its accuracy. You must also monitor and continue to measure error and/or consider multiple forecasting models to see if a certain method works better for your situation. There are so many forecasting models that I’ve only touched upon the surface here. Be sure that the forecasting efforts match the value it is creating. Artificial Intelligence is also opening up a brand new and promising way of forecasting. Keep in mind that businesses are usually not powerless against demand. There are many things that can be done with demand management (pricing, promotions, maintenance in the case of indirect spend, etc.) to proactively align supply with demand to deliver business objectives. As they say, the best way to predict the future is to create it.