Benchmarking Forecasting Models

The selection of a model plays a very important role in forecasting and one needs to know what kinds of models are available and which model is most suitable and where.

Abstract

The selection of a model plays a very important role in forecasting. Each dataset has a pattern of its own. Each model captures a specific data pattern. So for best results, it is important to use the right model for the right dataset. This raises a question of how to select the best model. To answer, one needs to know what kinds of models are available and which model is most suitable and where. Also, one needs to know some basics about models and modeling. There are basically three types of forecasting models: (1) Time Series, (2) Cause-and-Effect, and (3) Judgmental. TIME SERIES MODELS In Time Series modeling, we extrapolate the past data using one method or the other in search of the best statistical fit. Each Time Series model assumes that the past pattern will continue into the future. One of the Time Series models is the PercentChange method, which assumes that the average percentage increase/decrease in sales experienced in the past will continue into the future. If sales in the past increased on the average by 5%, then the forecast of the next month will be the sales of current period plus 5%. Time Series models are, by and large, easy to understand and easy to use. They generally work well for short-term forecasting because during that period the data pattern is less likely to change in a significant way. Times Series models include: (i) Averages including Simple and Moving, (ii) Simple Trend, (iii) Exponential Smoothing, (iv) Decomposition, and (v) Box Jenkins (also called Autoregressive Integrated Moving Average or ARIMA). Time Series models are also called univariate models because in these models we need only data of a series we wish to forecast. CAUSE-AND-EFFECT MODELS In Cause-and-Effect models, there is a cause (called a driver or independent variable) and there is an effect (called the dependent variable). If sales depend on the amount of money spent on advertising, then sales are the effect (dependent variable) and advertising is the cause (driver or independent variable). Here we determine the average relationships between the dependent (effect) and independent (cause) variables, and then use these relationships to project into the future. These models are used where there is a strong relationship between the causeand-effect variables and the relationship between them does not change significantly over time (or at least during the forecast period). Cause-and-Effect models include (i) Regression, (ii) Econometrics, and (iii) Neural Network. JUDGMENTAL MODELS Although judgment predominates in Judgmental models, these are by no means seat-of-the-pant models. There are set procedures that are used for forecasting. These models are often used where there are no historical data, as in the case of new products, if data do exist, they are no longer applicable, as in the case of fashion products (the pattern of one fashion product may not be the same as the other). Judgmental models are also used where the 80/20% rule applies (i.e., where a large percentage of sales comes from a few large customers); where the market is highly volatile; and where a forecast for a period far into the future has to be prepared (e.g., forecasting the price of crude oil 25 years from now). Among Judgmental models, models which are most often used in business are: (i) Analog, (ii) Delphi, (iii) Diffusion, (iv) PERT (Performance Evaluation Review Technique), and (v) Survey. In an Analog model, we look for an analog for a variable to be forecasted, and then use that parallelism as a basis for the forecast. For example, when television first came to the market, forecasters used radio as an analog to prepare sales forecasts for television. At present, businesses often use analog for the forecasts of line extensions of a product. In Delphi, a panel of experts is used to arrive at a forecast. In Diffusion, the forecaster follows the life cycle of a product to make a forecast. In PERT, the forecaster asks just one expert to give three estimates-pessimistic, optimistic, and most likely. …