Don't forecast

Data Science

Forecasting in human affairs is precarious

non stationarity impacts forecasting

connection to gödel? no system can be predicted from within the system itself?

—

From public policy newsletter, about radical uncertainty book:

Economics and social sciences deal with human behaviour. These social systems influence human behaviour and, in turn, are influenced by it. The sociologist Robert Merton called this phenomenon reflexivity. This makes forecasting difficult. In a time of radical uncertainty, human behaviour can be very different (non-stationary) from usual and starts impacting the system itself.

“Reflexivity undermines stationarity. This was the essence of ‘Goodhart’s Law’ – any business or government policy which assumed stationarity of social and economic relationships was likely to fail because its implementation would alter the behaviour of those affected and therefore destroy that stationarity.”

We don’t live in a ‘stationary’ world while all our social science models are built on that fundamental assumptions. Therefore, our ability to solve for or optimise the models that have been built for the ‘small world’ are bound to fail in the ‘big world’. These models should therefore be read like parables. There’s truth in what they offer which is important but you can’t take decisions in everyday life based on what you learn from them. Kay and King conclude:

Radical uncertainty and non-stationarity go hand in hand. There is no stable structure of the world about which we could learn from past experience and use to extrapolate future behaviour. We live in a world of incomplete markets in which there are simply no price signals to guide us back to an efficient equilibrium. There are times when expectations have a life of their own. As a result, the models used by central banks perform quite well when nothing very much is happening and fail dramatically when something big occurs – precisely the moment when the model might have something to offer beyond mere extrapolation of the past.