The model is not the business
A sportsbook makes money by taking the right amount of risk at the right price from the right customers. The model supplies the price; this lesson is about the rest. It is where quantitative desks that are excellent at pricing still lose, because a correct price with the wrong limits, or a correct price offered to the wrong customer, is a correct price the book is paying to publish.
Limits as a function of edge and uncertainty
The maximum stake the book accepts on a selection should depend on two things: how confident the book is in its price, and how much it is already exposed. Neither is a fixed number.
Confidence comes from the calibration work in lesson 3. In a competition where the model and the market agree and the calibration is tight, the book can take large bets because a bet that moves the price is unlikely to be informed. In a competition where the model is thin, the book should take small bets and move the price sharply on what it takes, because each bet may be information.
There is a formal version of this. The Kelly criterion, published by J. L. Kelly Jr. in the Bell System Technical Journal in July 1956, tells a bettor with an edge how much of their bankroll to stake to maximise long-run growth: the edge divided by the net odds, where the edge is the expected profit per unit staked and the net odds are the profit on a winning unit stake. In Edward Thorp's worked example, a 53% chance of winning at even money gives a stake of 6% of the bankroll. The book's problem is the inverse: it is offering a price and asking how much to accept from someone who might have the edge. The answer has the same shape. The acceptable stake scales with the book's estimated edge on the bet (its margin plus how much it trusts its price) and inversely with the variance of the outcome, and it is bounded by a risk budget for the event. Desks that implement limits as a Kelly-style function of estimated edge and exposure, rather than as a table of fixed numbers per market, take more where they are strong and less where they are weak, which is the whole point.
Liability and the event-level book
Exposure is managed per event in scenario space, as the previous lesson set out. The desk should be able to see, for every live event, the payout under every material outcome and the worst case. The management question is what to do when the worst case exceeds the event's risk budget.
The levers, in order of preference:
- Move the price. Shade toward the exposed outcome so the book takes more on the other side. This is the shape lever from lesson 1 and it costs nothing if customers keep betting.
- Cut limits on the exposed side. Slower to work, no cost.
- Hedge on an exchange or with another book. Costs the other book's margin, or on an exchange a commission on net winnings, and is only worth it when the exposure is genuinely uncomfortable, which for a well-diversified book is rare.
- Accept it. A book with thousands of events is diversified across them. A large exposure on one event that is within the overall risk budget is fine. Desks that hedge every uncomfortable position are paying away the diversification they own.
The risk budget itself is set above the desk: how much can the book lose on a single event, on a day, on a weekend, before it is a board conversation. Those numbers should exist, in writing, and the desk's job is to stay inside them at the lowest cost.
Customer pricing
Not every customer gets the same price, and this is the part of the business outsiders find hardest to accept. The mechanism is the per-customer weight from lesson 3, applied in two places: the customer's bets move the model probability in proportion to how informed they have proven to be, and the customer's limits are set by the same measure.
The commercially sound version of this is transparent about what it does. Sharp customers get lower limits on the markets where they are sharp and normal limits elsewhere; the book uses their information; the relationship is adversarial but stable. The commercially unsound version restricts anyone who wins, loses the information, and drives the informed money to competitors who then have better prices. Regulators have begun to ask books to justify restrictions. Massachusetts was the first US state to do so: since 1 June 2026 its rules have required operators to have procedures to give a limited bettor timely notice, with a specific explanation and the markets affected, and the Commission expects the notice within 48 hours and more than a generic reason. In that setting, "this customer beats our closing line by 3% over 400 bets" is a justification; "this customer is up" is not.
Bonus abuse, arbitrage and syndicate play are separate categories with their own detection, and they belong with the fraud team rather than the trading desk, though the desk supplies the signals.
Monitoring the model in production
A model that was good when it was deployed is not necessarily good now. The desk needs a monitoring surface that shows, per sport and competition:
- Rolling log loss and Brier score against the market baseline (both are strictly proper scoring rules, so neither rewards a model for misstating its probabilities)
- Calibration plot on the last N events
- Distribution of the model-market gap at open and at close
- CLV of the book's own opening prices against its closing prices
- Manual adjustment volume and the P&L attributed to it
- Stale-price bet cost in-play
- Parlay payout ratio against parlay implied probability
Each metric has a threshold and a response. The most important is the automatic one: when a model's performance degrades past its threshold, its weight in the blend falls toward zero and the market consensus carries the price until a human has looked. A desk that relies on someone noticing that the Danish handball model has drifted will find out in the P&L.
Backtesting without fooling yourself
Every model change is tested against history before it goes live. The traps are well known and still caught people:
Lookahead. Using information in the fit that would not have been available at the time of the price. Lineups, xG from the match being priced, closing lines. Every input must be timestamped and the backtest must only see what was known before kick-off.
Survivorship and selection. Testing on the competitions the desk chose because the model worked there. The same applies to model versions: when many configurations are tried and the best one is reported, the backtest is overfit to some degree, and a simple holdout does not account for how many were tried.
Overfitting to the test. Iterating a model against the same holdout until it wins. Every look leaks information from the holdout into the choices made, and testing ideas on the data used to explore them risks false discoveries even when the two use distinct subsets. Unless reuse is controlled by a method built for it, treat a holdout that has chosen between versions as training data and keep a fresh one for the final test.
Ignoring the market. A model that beats results is not the bar. The bar is beating the market consensus on the same events, because that is what the book can actually price against.
Ignoring limits. A backtest that assumes the book could have taken unlimited stakes at its model price is measuring something no book experiences. The realistic backtest simulates the flow the book would have seen and the limits it would have applied.
Governance: who can change what
A quantitative desk is a small group with a large lever. The controls that matter:
- Model changes go through review, shadow running and sign-off, with the evidence recorded.
- Manual price overrides are logged with a reason and reviewed weekly.
- Limit and margin schedules are owned by a named person and versioned.
- Risk budgets are set outside the desk and breaches are reported the same day.
- The desk's P&L is attributed: to model edge, to margin, to customer management, to luck. A desk that cannot separate these does not know what it is good at.
The through-line of this course is that pricing is a system, not a set of numbers: probability from models and the market, checked by calibration, kept coherent across markets and in-play, and turned into money by limits and customer management that reflect where the edge actually is. A desk that has all of that can price anything. A desk that has only the model is still guessing, just more precisely.