Last week, my research group dis-cussed Galit Shmueli’s paper “To explain or to pre-dict?”, Sta-tis-ti-cal Sci-ence, 25(3), 289–310. (See her web-site for fur-ther mate-ri-als.) This is a paper every-one doing sta-tis-tics and econo-met-rics should read as it helps to clar-ify a dis-tinc-tion that is often blurred. In the dis-cus-sion, the fol-low-ing issues were cov-ered amongst other things.

• The AIC is bet-ter suited to model selec-tion for pre-dic-tion as it is asymp-tot-i-cally equiv-a-lent to leave-​​one-​​out cross-​​validation in regres-sion, or one-​​step-​​cross-​​validation in time series. On the other hand, it might be argued that the BIC is bet-ter suited to model selec-tion for expla-na-tion, as it is consistent.
• P-​​values are asso-ci-ated with expla-na-tion, not pre-dic-tion. It makes lit-tle sense to use p-​​values to deter-mine the vari-ables in a model that is being used for pre-dic-tion. (There are prob-lems in using p-​​values for vari-able selec-tion in any con-text, but that is a dif-fer-ent issue.)
• Mul-ti-collinear-ity has a very dif-fer-ent impact if your goal is pre-dic-tion from when your goal is esti-ma-tion. When pre-dict-ing, mul-ti-collinear-ity is not really a prob-lem pro-vided the val-ues of your pre-dic-tors lie within the hyper-​​region of the pre-dic-tors used when esti-mat-ing the model.
• An ARIMA model has no explana-tory use, but is great at short-​​term prediction.
• How to han-dle miss-ing val-ues in regres-sion is dif-fer-ent in a pre-dic-tive con-text com-pared to an explana-tory con-text. For exam-ple, when build-ing an explana-tory model, we could just use all the data for which we have com-plete obser-va-tions (assum-ing there is no sys-tem-atic nature to the miss-ing-ness). But when pre-dict-ing, you need to be able to pre-dict using what-ever data you have. So you might have to build sev-eral mod-els, with dif-fer-ent num-bers of pre-dic-tors, to allow for dif-fer-ent vari-ables being missing.
• Many sta-tis-tics and econo-met-rics text-books fail to observe these dis-tinc-tions. In fact, a lot of sta-tis-ti-cians and econo-me-tri-cians are trained only in the expla-na-tion par-a-digm, with pre-dic-tion an after-thought. That is unfor-tu-nate as most applied work these days requires pre-dic-tive mod-el-ling, rather than explana-tory modelling.