Select Page
Report an error

11. Forecasting

Types of Forecasts:  Quantitative and Qualitative

When we forecast, we generate predictions for future values, using known values from previous observations, along with other information that can either come from within the data (such as observable patterns), or from external sources (perhaps we have reason to believe that some other variable will impact the quantity that we wish to predict).  Forecasting can have many applications in marketing analytics.  Here are just a few:

  • If Lobster Land managers can predict the number of expected lobster roll orders this week, they can optimize the size of their purchase order from their seafood wholesaler;
  • If Lobster Land’s finance team can predict the total number of season passholders who will renew their passes for next year, they can use that number to help formulate a budget for offseason capital expenditures;
  • If Lobster Land’s operational team can forecast daily ridership demand for the Lobster Claw and the Ferris Bueller, they can determine optimal staffing allocations for those rides prior to the morning “huddle” among the managers and the hourly employees.  

Forecasts may be quantitative, they may be qualitative, or they may be a blend of the two types.  

Quantitative forecasts rely on numeric inputs, and their underlying patterns and distributions, to predict future values in a series.  They will be our main focus for this chapter.  We will start out with some very simple (yet sometimes quite effective) techniques, before moving on to more complicated model types.  

Qualitative forecasts, on the other hand, rely on forecasters’ domain knowledge, subjective judgment, and intuition.  

Why use qualitative forecasts?  At times, quantitative data may be unavailable.  In other instances, the problem might not be suitable for numeric analysis.  Imagine asking Cold War researchers, in the mid-1980s, “Do you think the Berlin Wall will come down within the next 10 years?”  This is fundamentally different from, say, predicting Walmart’s online sales for next quarter – in the latter case, there are plenty of previous, known values to learn from.  

Qualitative forecasts are subject to many forms of human bias.  Among these are:

  • Overconfidence bias.  This occurs frequently in the world of investing.  An investor who makes one great decision, or who earns outsized returns for a certain time period, may begin to feel invincible.  
  • Anchoring bias.  Anchoring happens whenever we base our forecast for a future period on another, known value.  While this is often a good starting point for any forecast, it ignores changes that may have occurred between periods.  
  • Confirmation bias.  This occurs when people selectively filter the evidence that they hear in order to confirm an existing belief.  
  • Groupthink bias.  This occurs when forecasters allow conventional wisdom to reduce the objectivity of their predictions.  This can even occur within organizations, when the leader announces his or her opinion first, before soliciting others for their opinions.  

Even with those biases in mind, we caution you to never fall into the trap of believing that a quantitative forecast must be better than a qualitative one, simply because it’s based on numeric inputs.  As the old saying goes, “Garbage In, Garbage Out.”  In fact, the more familiar you become with quantitative forecasting models, the more skeptical you should become about any such model.  

An example of a blended forecast could be a model that attempts to predict whether the U.S. economy will go into a recession within the next two quarters.  This model could use numeric inputs such as trends in employment, labor force participation, and large capital expenditures from businesses.  It could then also use qualitative inputs based on the subjective opinions of experts – for instance, one input could simply be based on economists’ answers to Yes/No questions such as: “Do you think that U.S. companies’ payrolls are likely to contract within the next six months?” or “Do you think that the Federal Reserve is likely to raise the Federal Funds Rate for borrowing within the next six months?”  

In Chapter 9, we learned about ensemble learning models.  In an ensemble model, a prediction is generated based on the “votes” of multiple models, none of which is identical to the others.  Ensemble modeling can also be performed with forecasting, and it can work the same way.  A modeler could generate several unique forecasts, and then average their results together.