Why the outputs of models should be interpreted as predictions

Abstract

Physicists and engineers are accustomed to seeing the output of a computational model as representing a prediction. In this context, predictions are understood to be informed approximations of how a system's dynamics may evolve as a result of the modelled assumptions. In what is generally called inverse modelling or optimization, applied mathematicians employ the ability of models to reproduce observed behaviours in order to deduce causal relations or to retrodict the system's past history. However, in the modelling of ecological and social processes, scientists (including modellers themselves) tend to be much more pessimistic concerning the ability of models to provide reliable predictions, Because of this sceptcism prediction is sometimes explicitly excluded from the list of useful model purposes (1).The four principle reasons for this scepticism are: a) computational models have a very poor prediction track record; b) most model predictions are not testable because of their conditional nature; c) despite the appearance of objectivity, model outcomes reflect the modelers' subjective beliefs and assumptions and d) the view that some scientific activities,including computational modelling of social and ecoological phenomena are not designed, and should be expected to, provide predictions. While we acknowledge that the modelling of physical vs ecological and human processes are different (2-4), we suggest that the output of any type of model which is employed as part of a decision-making process should be interpreted as a prediction. Our claim is based on the following reasoning. First, prediction should not be understood as a forecast of a precise event, instead it should be understood as an estimation of a probability distribution which provides bounds on the likelihood of sets of future events. Second, such predictions are indispensible for decision making, since they provide the basis upon which available alternative options are evaluated and chosen. Third, the predictive power of experts is known to be less reliable in certain contexts, than numerical models. Given that prediction is necessary for non-arbitrary decision making, it is useful to refocusing the question from whether models provide an accurate prediction to whether computational model can outperforms humans as predictors. Fourth, the ability to compare prediction vs observation is at the core of the scientific method and dismissing the predictive capacity of models prevents blocks the possibility of assessing the relative scientific merits of distinct models. Fifth, disregarding the predictive power of computational models prevents their use in inverse modelling. Other familiar uses of computational modelling, like deducing causal relations and past system behaviour are also logically denied. Sixth, further commonly recognized purposes of numerical modelling, like leaning and communication, also rely on the ability of models to predict within acceptable limits: it is pointless to learn from a mistaken teacher. Seventh, acknowledging the role of prediction in assessing a model's scientific validity and its impact on decision making forces modellers to accept the responsibility of providing all the necessary details of the model so that the reliability of such prediction can be estimated. Below we expand on these points. In an attempt to reconcile different views on models, predictions and their merit in decision making, we conclude by providing an alternative interpretation of computational models, according to which models can be understood as an extension of native cognitive capacities.