Covering Predictive Analytics in the Classroom

Graph showing Interest over time for Business Analytics and Predictive Analysis

There has been a corresponding growth in the number of courses offered. In the US now almost all the top universities offer masters’ level programmes in analytics, although their designation varies and often includes data science. In the UK major providers of operations(al) research masters’ programmes have changed their titles if not their content to business analytics. Mortenson (2019) has attempted to separate out the different types of programme and their content in the UK. In the US we have undertaken an analysis of the top 50 programmes in analytics. While each programme has its own individual features and it is often hard to establish what is in fact taught, pretty well all the programmes have a core statistics course, perhaps with some limited computer programming, such as R. Moving on to more advanced topics, we can pick apart what is meant by predictive analytics. 

The Wikipedia definition is: “Predictive analytics is an area of statistics that deals with extracting information from data and using it to predict trends and behaviour patterns. Other definitions are similar and include aggregate and individual behaviour as well as events. Keywords include ‘Artificial intelligence/ Machine learning’, ‘Classification and clustering’, ‘Time series analysis’, ‘Econometrics’ and ‘Forecasting’. In addition, applications areas such as ‘Operations’, ‘Revenue management’ and ‘Retailing’ might well include relevant material. An examination of the top masters’ programmes only permits an impressionistic view. Machine learning methods such as neural nets, probably associated with classification and clustering with cross sectional data predominate, followed by time series analysis. Although regression analysis is widely recognized as a standard tool, it appears to be used primarily for cross-sectional analysis, perhaps a reflection of the obsession with “big data”. By contrast, econometrics and forecasting are seldom mentioned. 

Are the time series courses sufficient to give students a full understanding of forecasting? The emphasis is typically on model building: forecasting is only mentioned in passing. Yet, forecasting is an activity that most organizations necessarily carry out: this will include sales forecasting, cash flows, cost projections, and the analysis of marketing campaigns as well as long term ‘scenarios’ that require the use of managerial judgment. The field is far more all-embracing than ‘time series analysis’ or even ‘econometrics’ and includes certain key elements which would not naturally be included in the more standard university courses on offer. Without any attempt at completeness we suggest that a comprehensive course of forecasting should include:

1. The definition of the forecasting problem, including the decision context, the forecast horizon, the information available to the forecaster, and the evaluation of the forecast’s validity.

2. Basic forecasting methods such as exponential smoothing: this approach is widely used and yet the topic is often excluded from time series modules.

3. Regression model building: while basic statistics courses and econometrics courses discuss the mechanics of model estimation, the process by which a valid model is built and tested is all too often ignored.

4. Machine learning methods as they apply to time series: these are much neglected and the problems of avoiding overfitting often ignored.

5. The particular problems faced by companies relating to operations, finance and marketing. The limited time available in an applications course means that all too often incorrect recommendations are offered.

6. Judgment in forecasting: this is perhaps the most neglected area and yet surveys have shown it to be the most prevalent in practice, both for generating forecasts and adjusting those developed by quantitative methods.

7. Finally, perhaps most important, the role of ‘competitions’ in selecting an appropriate method: it is at the heart of data science or should be (Donoho, 2017). A fundamental issue is how to measure predictive accuracy and the value of improved methods.

So despite the broad definition of predictive analytics, it seems that in most courses there has been an overall focus on cross-sectional modelling and out-of-date notions of how to develop an appropriate ‘predictive’ time series model. Our new book (Ord et al., 2017- Principles of Business Forecasting) aims to explain and illustrate all these important features of forecasters’ problems as well as explaining time series analysis and regression in ways that emphasize their practical relevance. For the practicing forecaster, it also includes many worked examples so that it operates as a manual of how these methods can be implemented with its R programs.

Why this neglect of forecasting, particularly in North America? We speculate this arises because of its multidisciplinary nature which means the topic lacks a departmental sponsor. The fundamental developments in forecasting research have arisen from Brown (in operations management), Box and Jenkins (in statistics), Zellner, Granger and Hendry (econometrics) and Kahneman and Tversky (in psychology). All these disciplines have a crucial role to play in predictive analytics. We recommend that all predictive analytics programmes, both undergraduate and masters, institute a forecasting course. Their omission short-changes the students and leaves courses in ‘decision analytics’ without the necessary foundations. From our experience, students very much enjoy engaging with forecasting, not least because they can see the practical relevance and career value. 

Back to News