How to bet on football with a model

A model that says the home side wins 54% of the time has told you nothing you can act on. It becomes actionable only in combination with three other things, and the order matters.

#One: a price to disagree with

54% is a bet if the market is charging for 48% and not a bet if the market is charging for 56%. The number alone is inert. This sounds obvious and it is the single most common failure: people build a model, look at its output, and back the side it likes, which is backing favourites at whatever price is going.

The comparison is the bet, which is what an edge actually is.

#Two: an honest account of the error

Your 54% has error bars around it, and if they are five points wide then a market at 51% is not a disagreement, it is noise. A model with no stated uncertainty produces a stream of bets that are indistinguishable from a model with a small edge, and both look identical for months.

#Three: a stake that survives being wrong

Even a correct 54% loses 46% of the time, and it will do so in runs of eight and ten. If those runs end the bankroll then the model's quality never becomes relevant, which is the entire argument of staking without going broke.

#What a football model can realistically do

Beating a major league's match odds is very hard; those markets are the most heavily traded in the sport. The realistic targets are derivatives priced off the headline number, smaller leagues where nobody is paying attention, and situations where information arrives faster than the market absorbs it.

#What this does not tell you

Nothing here says any particular model has an edge, and a model that fits the last three seasons beautifully has usually fitted the noise in them. The test is out of sample, against the closing price, over a period long enough to be boring.