Part III: PREDICTION

Chapter 13: A Framework for Prediction

This chapter introduces a framework for prediction. We discuss the distinction between various types of prediction, such as quantitative predictions, probability predictions, and classification, and we focus on the first of these. We introduce point prediction versus interval prediction and we discuss the components of the prediction error. The main focus of this chapter is how to find the best prediction model, using observations in the original data, that will likely produce the best fit (smallest prediction error) in the live data. We introduce loss functions in general, and mean squared error (MSE) and its square root (RMSE) in particular, to evaluate predictions. We discuss three ways of finding the best predictor model: using all data and the Bayesian Information Criterion (BIC) as the measure of fit, using training–test splitting of the data, and using k-fold cross-validation, which is an improvement on the training–test split. We discuss how to assess and, if possible, improve the external validity of predictions. We close the chapter by discussing what machine learning means.

Chapter 14: Model Building for Prediction

This chapter discusses how to build regression models for prediction and how to evaluate the predictions they produce. With respect to model building, we discuss whether and when it’s a good idea to take logs of the y variable and what to do with such a prediction, as well as how to select variables out of a large pool of candidate x variables, and how to decide on their functional forms and including their interactions. We introduce LASSO, an algorithm that can help with all that. With respect to evaluating predictions, we discuss why we need a holdout sample. We close this chapter with a discussion on the additional opportunities and challenges Big Data brings for predictive analytics.

Chapter 15: Regression Trees

This chapter introduces the regression tree, an alternative to linear regression for prediction purposes that can find the most important predictor variables and their interactions and can approximate any functional form automatically. Regression trees split the data into small bins (subsamples) by the value of the x variables. For a quantitative y, they use the average y value in those small sets to predict ˆy. We introduce the regression tree model and the most widely used algorithm to build a regression tree model. Somewhat confusingly, both the model and the algorithm are called CART (for classification and regression trees), but we reserve this name for the algorithm. We show that a regression tree is an intuitively appealing method to model nonlinearities and interactions among the x variables, but it is rarely used for prediction in itself because it is prone to overfit the original data. Instead, the regression tree forms the basic element of very powerful prediction methods that we’ll cover in the next chapter.

Chapter 16: Random Forest and Boosting

This chapter introduces two ensemble methods based on regression trees: the random forest and boosting. We start by introducing the main idea of ensemble methods: combining results from many imperfect models can lead to a much better prediction than a single model that we try to build to perfection. Of the two methods, we discuss the random forest (RF) in more detail. The random forest is perhaps the most frequently used method to predict a quantitative y variable, both because of its excellent predictive performance and because it is relatively simple to use. Ensemble methods are black box models, because their results do not help understand the underlying patterns of association between y and the x variables. We discuss some diagnostic tools that can help with that: variable importance plots, partial dependence plots, and examining the quality of predictions in subgroups. Finally, we briefly introduce the idea of boosting, an alternative approach to make predictions based on an ensemble of regression trees. There are various boosting methods, and they can produce even better predictions, but their use requires more expertise. We illustrate the power of boosting through the performance of the gradient boosting machine (GBM) method.

Chapter 17: Probability Prediction and Classification

This chapter introduces the framework and methods of probability prediction and classification analysis for binary y variables. Probability prediction means predicting the probability that y = 1, with the help of the predictor variables. Classification means predicting the binary y variable itself, with the help of the predictor variables: putting each observation in one of the y categories, also called classes. We build on what we know about probability models and the basics of probability prediction from Chapter 11. In this chapter, we put that into the framework of predictive analytics to arrive at the best probability model for prediction purposes and to evaluate its performance. We then discuss how we can turn probability predictions into classification with the help of a classification threshold and how we should use a loss function to find the optimal threshold. We discuss how to evaluate a classification making use of a confusion table and expected loss. We introduce the ROC curve, which illustrates the trade-off of selecting different classification threshold values. We discuss how we can use random forest based on classification trees. Finally, we note the potential issues with the probability prediction and classification of rare events.

Chapter 18: Forecasting from Time Series Data

This chapter discusses forecasting: prediction from time series data for one or more time periods in the future. The focus of this chapter is forecasting future values of one variable, by making use of past values of the same variable, and possibly other variables, too. We build on what we learned about time series regressions in Chapter 12. We start with forecasts with a long horizon, which means many time periods into the future. Such forecasts use information on trends, seasonality, and other long-term features of the time series. We then turn to short-horizon forecasts that forecast y for a few time periods ahead. These forecasts make use of serial correlation of the time series of y besides those long-term features. We introduce autoregression (AR) and ARIMA models, which capture the patterns of serial correlation and can use it for short-horizon forecasting. We then turn to using other variables in forecasting, and introduce vector autoregression (VAR) models that help in forecasting future values of those x variables that we can use to forecast y. We discuss how to carry out cross-validation in forecasting and the specific challenges and opportunities the time series nature of our data provide for assessing external validity.