What a football model teaches you about forecasting markets
In five days the World Cup kicks off. I built a model to forecast it. The model is not the point. The point is that football, unlike markets, grades you in public and on a deadline.
Most forecasting lives in a comfortable fog. You make a call, the world moves, and by the time the outcome arrives the question has changed enough that nobody checks. Markets are the worst offender. A view on equities in June is unfalsifiable by December because ten other things happened in between. Football has no such mercy. The whistle blows, the score is the score, and four weeks from now everyone can see whether the model was right.
So I treated the World Cup as a stress test of the same discipline I apply everywhere else.
The approach, in outline
The model is an ensemble. It combines more than one independent statistical engine, each estimating team strength and match outcomes a different way, then blends them. It layers in external strength signals beyond raw results, and it recalibrates for the things simple models get wrong, draws chief among them. On top of that sits a Monte Carlo simulation: the tournament is played forward 10,001 times over the new 48-team, 104-match format, and the championship probabilities are the frequencies that fall out.
I am keeping the internals to myself. The value of a model is not the idea, which is freely available in any sports-analytics paper, it is the calibration, the choices, and the hours of getting the details right. What I will share is the output and, more importantly, the evidence that it works.
The discipline that actually matters
Anyone can produce a forecast. The question is whether you can produce one that beats doing nothing. So the model is validated the way I validate anything before I trust it.
Walk-forward backtest, no peeking: every prediction is made using only data available before that match. Over the last 12 months of international football, 997 matches, the model scored a mean Ranked Probability Score of 0.165. An uninformed forecast scores 0.278. A perfect one scores 0. Outcome accuracy came in at 50.8 percent against a 33.3 percent baseline for blind three-way guessing. Across five recent major tournaments it beat both naive benchmarks and every single component model in the battery.
That last point is the one I care about most. A model has to earn its complexity. If the elaborate version cannot beat a simple rule, the elaborate version is vanity. This one clears the bar, but only modestly, and I will say so plainly: the edge is real and it is small.
What it says
Spain, champions, at 19.8 percent. Then France at 12.6, Argentina at 12.2, Brazil and England level at 8.8. The top four hold 53 percent of the title probability between them. The most likely final is Spain against Argentina, with England and France meeting in the third-place game.
Read those numbers correctly. Spain at 19.8 percent means Spain loses this tournament four times out of five. The single most likely bracket, the one where every favorite advances exactly as projected, has a joint probability close to zero. That is not a flaw in the model. It is the truth about football, and the same truth holds for markets. The headline call is the least interesting number on the page. The distribution is the forecast.
Why this connects to the day job
Three habits carry directly from this exercise into how I think about markets.
First, ensemble over conviction. Two independent methods that disagree tell you more than one method you happen to like. Where they agree, lean in. Where they diverge, the gap is information, not noise to be smoothed away.
First principles on uncertainty. A probability is a statement about long-run frequency, not a prediction. Spain at 19.8 percent and a portfolio position sized to a 20 percent base rate are the same kind of claim. Treat them the same way.
And backtest honestly or do not bother. The temptation in every model is to let a little future information leak into the fit and admire the result. The whole value of the football version is that the leak gets exposed in public, on a fixed date, with no second question to hide behind. If a process cannot survive that, it should not be running your money either.
What happens next
This is version one. I will keep iterating up to kickoff and through the tournament as results land. The model will be wrong about specific matches, often. The test is not whether Spain lifts the trophy. The test is whether, across 104 matches, the probabilities turn out to be calibrated. That is the only thing worth measuring, and unlike most of what I do, you will all get to watch it happen.
from Risk Premium https://ift.tt/QabToZv
via IFTTT
No comments:
Post a Comment