GuidesPokerSports BettingStrategy
How to Build a Sports Betting Model From Scratch
A gentleman from Hamburg. His cheque cleared at four in the morning. His model cleared every projection we had. An examination of how serious bettors construct the thing itself.
Words by Carol Feldman4 min read
In the high-limit room at Monte Carlo, we see certain kinds of money come through. The hasty money, frightened money, money trying to prove something. But every few years a gentleman arrives with calculations in his eyes and a model that works. They are rare because building a functional model requires something most bettors cannot supply: the willingness to watch data instead of watching the action.
A model is, at its simplest, a system for turning data into a probability. You collect outcomes from previous games. You identify variables that seem to predict future outcomes. You build a formula that weights those variables. Then you test it against outcomes it has never seen. If it works, you use it. If it doesn't, you adjust or you abandon it. The gentleman from Hamburg brought a model that had been refined over five seasons. He understood what most bettors never grasp: a good model is boring to use. It tells you when to bet and when not to bet, and it does not care what you want to do.
The Data Problem
Most people begin with too much data and too little sense of what they're trying to predict. They collect score differentials, home-field advantage, weather, injury reports, player salaries, phase of the moon. The model becomes ungainly. It finds patterns in noise. It overfits to the past and breaks when the future arrives.
The best models I've watched were lean. One gentleman who played high-limit poker and sports betting with equal seriousness built his entire system on four inputs: pace of play, turnover differential, shooting efficiency, and rest days. He collected twenty years of NBA data. Every season, he tested his model on games it had not seen. He did not bet every game. He bet when his model said the line was off by more than 3 percentage points. He went small. He compounded slowly. After five years of this he had a real edge.
The work is not glamorous. You sit with spreadsheets. You check for errors. You discover you miscoded a formula and have to start again. You build the thing and it fails and you figure out why. This is the part where most people quit. They want to be wrong in a way that feels like an insight. Instead they find out they were wrong because they didn't understand the data well enough to know what they were measuring.
Most people begin with too much data and too little sense of what they're trying to predict.
Testing the Model
Once you have data and variables, you test backwards through time. Take the first two seasons. Build your formula. See if it predicted the third season correctly. Roll forward. Use seasons three and four to predict season five. The Hamburg gentleman did this for five seasons each, twelve different tests, testing on data the model had never encountered.
He found his edge was about 2.8 percentage points on average, which sounds small until you realize that professional lines are only accurate to within 2 to 3 percentage points overall. His advantage was marginal but real. Crucially, he also discovered when his edge disappeared. The model worked fine on games with spreads between two and five points. At the extreme ends of the spread spectrum, it had no predictive power. So he did not bet on those games. He took money when the model said to take it and stayed out of the rest.
This is where discipline enters. Most bettors, given such a model, would find reasons to bet on everything. They'd say the model was wrong about the extremes, or they had a hunch, or they'd seen something that changed the picture. The gentleman from Hamburg did not. His discipline was boring and functional. He bet small. He tracked everything. He accepted that variance existed and could wipe him out for a season. He stayed in the game.
The Real Edge
What distinguished his model from the thousands of failed models was that he tested it honestly. He did not adjust for bad outcomes. He did not add variables after seeing the results. He did not convince himself that a losing streak meant the model was right and the market hadn't caught up yet. When the model failed on certain kinds of games, he accepted that failure instead of double-betting to prove himself right.
Building a model is not about intelligence. It's about accepting that you are probably wrong and designing a system to find out how wrong you are, quickly and cheaply. Once you know, you adjust or you stop. Most people cannot do this. They attach their ego to the model. They want to be the person who beat the house. They don't want to be the person who built a system that turned a 2.8% edge into retirement in ten years of patient, small-bet work.
The gentleman from Hamburg was one of the few who understood the difference.