Using Data for Quantitative Forecasting
Forecasting is the idea of making good estimates for future performance and potential results. Forecasting falls into two subjects, Qualitative (an opinion of a person or group of people) and Quantitative (a judgment based on existing data). Qualitative could be called “best guess” but it is not accurate or most often is not good. In other words, qualitative is subjective.
Quantitative forecasts are usually something that can be measured and also assumes the same general patterns will continue. Since quanitative forecasts are based on facts and existing data, it is ultimately objective and without ambiguity. We can then break quantitative forecasts into two smaller subjects: Casual and Time Series.
Quantitative Forecast Types - Casual and Time Series
Casual is the idea of “cause and effect”, or that the data is influenced by one or more external reasons. For instance, if we have a two-for-one sale on widgets, there could be an increase in how many widgets are purchased. If we lower the price on a widget, we might also see an “uptick” in trends. If we have done things like this in the past (promotions, price changes), we can use that data to predict what we think will happen in the future.
Using historical data, we can use the Time Series methods to assume that patterns will generally remain the same in the future, and the shorter the time “frame” (three years versus five years, for example) will give a more accurate projection of the future. Time Series can also help to identify what factors influence changes in trends. We can use existing equations and algorithms, also called Models, to handle projections and forecasts.
So where can we start? We can use the Holt-Winters method to handle our data. Why? The Holt-Winters method is the most commonly used forecasting model in business. It’s so good it’s still in use after 55+ years.
What is a Holt-Winters Forecast?
Holt-Winters is the idea of using a set of data (best when it's a group rather than individual variables) to estimate the current levels (normalized by removing noise and seasonal data), the current trend and the “time” index of those sets of data (quarters, months and years). The equation looks like [Level+ (2*Trend)] * TimeIndex = forecasted value.
That’s forecasting…when we turn the forecasted value into lines, we need to figure out how to (adjust\adapt) weigh our values to consider older data versus newer data. Holt-Winters assigns a greater weight to more recent values, meaning we “smooth” out our trend line and adapting for outliers using ratios (estimated and real). That’s where the “simple smoothing” differs from Holt-Winters. It doesn’t consider the weights of the values in relation to the age of the value.
The standard version of a forecast assumes we have both the time series and the trend. When we are missing the time or the trends, we use the appropriate forecasting model. When we just want the average (which has value in predicting what the changes are, but is terribly inaccurate) to spitball a future value, we use Simple Smoothing for the trends. When we don’t really want to use the Holt-Winters method and just want to smooth out a trend and get a slightly more accurate prediction, we would use Exponential Smoothing, which will make a best guess at future performance based on past patterns.
Irrespective of the business vertical or the functions of a role in a company, accurate future estimates are needed. Qualitative forecasting (“a rough guess”, “speculated” or “a hunch”) can lead to losses from overestimating or underestimating the future. How many sales are lost if there isn’t enough product to meet demand? How much will it cost to store excess product? How much in sales could be lost by offering discounts to remove the excess product? How much of an impact would employee morale take if unrealistic goals are set? How dissatisfied would investors be if predicted profits aren’t met? These real effects of subjective forecasting shows why data driven forecasting is a necessary part of every business.