On Models

I have tried, and admittedly occasionally failed, to avoid these posts becoming too focussed on my particular annoyances and frustrations, a risk in what can be a self-indulgent form of communication. Nevertheless, I feel I must, for professional as well as personal reasons, unburden myself in relation to an irritation that I feel with the abuse of the term 'model'. This might seem a strange thing to get hot under the collar about, but modelling is the backbone of engineering and a tool of immense power and reach, to see it traduced seems to me to be an act of intellectual vandalism.

Let me be clear then - management gurus, pay attention - a PowerPoint slide with a pretty diagram, is not a model - it might be (or aspire to be) a framework, a tool for thought, or whatever, but it is not a model. Similarly - computer scientists, pay attention - just representing something using mathematical symbols, does not make it a model, if before you started to represent it you were confused, afterwards you are merely confused in maths. Dumping all the data you have into an arbitrarily constructed simulation program is not modelling - physicists, yes, you - the only things you have added are bugs.

A model is a disciplined simplification, there must first be a simplification - an abstraction - and this must be principled, in other words detail must be removed (from a collection of real-world observations, a model requires empirical grounding) systematically and in accordance with a set of consistent principles. The basis for this simplification must itself be amenable to formal analysis. Your model must have predictive power, if you cannot analyse the representation to obtain knowledge you did not possess at the outset, it is not a model. Ordering your thoughts in some sort of structured manner is important and worthwhile, but this process is not modelling and the result is not a model. In a model the choice of representation is of the essence. The modelling language must chosen with care to yield analytic traction and to make the process of translation from observations to model easier. The validation path from model, through predictions to observation must be navigable. The path from model back to the structure of the observations must also be navigable. Only then do you have a model. There, thank you, I feel better now.