Algorithm Aversion: On Models and The Donald

Whatever your politics, one can’t help but be fascinated by the Trump phenomenon. One interesting aspect is how pollsters got Trump wrong. We can get insight into that mistake from Nate Silver at fivethirtyeight. To his immense credit, Silver admits he got Trump all wrong, and went back to have a look at the why and the how…. it boils down to a case of algorithmic aversion (see here and here).

Silver’s approach to forecasting, which has been extremely successful across a wide range of subjects, is based on aggregating data from multiple polls to build as big a sample as possible. It is a mechanical, quantitative, model based, statistical method. The model spits out an answer, and that is forecast. But in the Trump case, he over-rode the model. He had a bias, and in this instance he applied it to his forecast, and voila, a bad outcome. Even Silver, a devoted algorithmist if there ever was one, fell into the trap. He did not trust the model, even though it had been right so many times before.

Continue reading

Bullard, Regime Switching and Trend Following

St Louis Fed President James Bullard has released a paper detailing a revised approach to economic forecasting. It’s a very smart way of looking at the world – read it here. Briefly, he is saying that the current way of viewing the world as converging to a single state is no longer useful and instead should be thought of as a set of possible regimes the economy could visit, with the regimes being generally persistent, requiring different monetary policy responses, and switches between regimes as not being forecastable. In his submission to the FOMCs quarterly economic projections, he declines to provide a forecast for the ‘Long Run’, as it is outside his model projection range. His low projection of the Fed Funds rate over the coming years reflects his view that the present regime has a low neutral real interest rate, a switch to a higher regime is unforecastable. If it were to happen, it would cause a change to many variables – policy would not reflect a gradual shift to a single state, but would have switched regimes.

This approach to forecasting was pioneered by James Hamilton. The math is pretty complex (lots of markov processes, etc.), but here is a simple way to look at it. Suppose we have two possible states in the world, the bull state and the bear state. The variable that determines the state is unobservable, and since you can’t see it, you can’t forecast it. Suppose in the bear state that the daily returns to an asset, like the stock market, are selected from a normal return distribution with a negative mean. Conversely, in the bull state, the mean is positive. If the state variable is pointing at bear, the trend will be down, if bull, the trend will be up. The trendiness of a market is determined by how likely we are to remain in the current state. For example, if the probability of jumping from one state to another is 5%, trends are more likely to persist than if the probability were 20%. What causes the state variable to jump is unknown, as Bullard describes.

Continue reading