Kelly Criterion Across Domains 2026 — Horse Racing, Poker, Polymarket

Kelly Criterion across horse racing, poker, and prediction markets — domain comparison collage — Photo via Picsum

The Kelly Criterion is the same formula whether you're at the track, at a poker table, or buying YES contracts on Polymarket. The math is identical. The operating reality of each domain is wildly different, and that reality changes how the formula gets applied. This post walks through the three big betting domains where Kelly is most useful in 2026, what works, what breaks, and what pros actually do in each.

Horse racing: pari-mutuel wrinkles

Horse racing was where the modern theory of bet sizing was forged. Bill Benter built one of the most successful gambling syndicates in history applying Kelly to Hong Kong horse racing in the 1990s. He turned a few thousand dollars into reportedly hundreds of millions. His version of Kelly was anything but textbook.

The wrinkle in racing is pari-mutuel pools: the odds aren't fixed when you bet. They drift as more money flows in. Your sizing decision changes the odds — especially on smaller pools. A naive Kelly bet that assumes static odds will systematically over-bet because the act of placing the bet erodes its own edge.

The pro adjustment: model the final odds after your bet has been absorbed into the pool. Iterate. Bigger bets push your effective odds toward the average; smaller bets keep your edge. The fixed-point math is similar to walk-the-book on a Polymarket order book.

Other racing-specific Kelly adjustments:

Multiple horses in one race are correlated — only one can win
Track-takeout (the rake) can be 15-25% — your edge has to clear that before Kelly even applies
Late-money information: closing odds are often more informative than mid-day odds

Poker: estimating opponents, not nature

Poker Kelly is fundamentally different from racing or prediction markets because your edge isn't against nature, it's against humans. Your win probability depends on your read of a specific opponent in a specific spot.

Three twists in poker:

Edge is heterogeneous. A weak opponent might give you 10pp of edge; a strong one gives you 1pp. Kelly sizing should reflect that — bigger bets in spots against weaker opposition.
Bet size is constrained by table. You can't bet $5,000 on a $50 buy-in tournament. Sometimes Kelly recommends a position you literally can't take.
Variance is brutal. Even an A-game player can be down 20 buy-ins in a swingy month. Kelly applied to poker requires fractional Kelly with strong discipline — full Kelly destroys careers.

Pro poker players don't typically run Kelly bet-by-bet. They size their session and game selection with Kelly logic: which game has the best edge given my bankroll, with what fraction should I sit. The classical example is the Ed Miller / Mason Malmuth formula for cash game bankroll requirements, which is essentially Kelly's drawdown math applied to game stakes.

Prediction markets: textbook home turf

Polymarket and similar venues are where Kelly fits best with the least adjustment. The reasons:

Binary outcomes mapping cleanly to the textbook formula
Continuous prices that already represent probabilities — no odds conversion needed
Independent bets across unrelated markets (mostly)
Persistent prices you can size against without pari-mutuel drift

The main wrinkle is order book depth — your stake walks the book and raises your effective price. That's the walk-the-book adjustment, mathematically similar to the pari-mutuel adjustment in racing but easier to model because the book is visible at the time of the bet.

Long resolution windows are the other constraint. A 6-month resolution locks your capital, so Kelly should be sized off your available bankroll (not total Polymarket account balance) to reflect what you can actually deploy concurrently.

What's identical across all three domains

The math, of course. The Kelly fraction f = (b·p − q) / b is invariant. But also:

Fractional Kelly is universally correct. Quarter Kelly is the right default in horse racing, poker, and prediction markets alike. Estimate noise is everywhere; full Kelly is for textbook problems.
Robust shrinkage applies everywhere. Treat your point estimate as a Beta posterior, size off a credible bound. The mechanic is the same regardless of whether your edge comes from horse form analysis, poker hand reading, or political modeling.
Correlation kills you everywhere. Multiple horses in one race, multiple hands at one table, multiple markets on one geopolitical event — same problem in three forms. Cap your cluster exposure.

What's domain-specific

Aspect	Horse racing	Poker	Prediction markets
Edge source	Form analysis, track conditions	Opponent reads, hand math	Probability modeling, news
Price impact	Pari-mutuel drift	None per-hand	Walk-the-book
Resolution lag	Minutes	Minutes per hand	Days to months
Effective rake	15-25% takeout	Casino rake / tournament fee	Spread + ~2% maker fees
Best Kelly fraction	0.25 (high noise)	0.10-0.25 (high variance)	0.25 standard

What pros do in each

Horse racing pros run quantitative models against historical race data, apply Kelly with pari-mutuel adjustment, and bet at quarter Kelly with hard caps on individual race exposure.
Poker pros rarely apply Kelly bet-by-bet. They use Kelly logic for stakes selection and bankroll management — typically requiring 30-50 buy-ins for cash games and 100+ for tournaments.
Prediction market pros use Kelly closest to textbook, with walk-the-book sizing, robust shrinkage on estimates, and correlation-aware cluster caps. Quarter Kelly is the default; some go to half on highly calibrated edges.

Pros and cons of cross-domain Kelly

Pros

Same math, transferable intuition across betting domains
Forces discipline around computing edge per opportunity
Robust shrinkage and fractional Kelly principles work everywhere

Cons

Domain-specific frictions (rake, drift, depth) can each kill edge if ignored
Correlation handling is non-trivial in every domain
Kelly is a recommendation, not a guarantee — every domain still has variance

Frequently asked questions

Which betting domain is friendliest to Kelly?

Prediction markets like Polymarket. Binary outcomes, continuous prices already in probability form, and visible order books make the formula's assumptions cleanest to satisfy.

Why don't poker pros use Kelly bet-by-bet?

Because their per-hand edge estimates are high-variance and they're constrained by table structure (you can't bet exactly Kelly's recommended stake when the action is forced into discrete options). Kelly logic shows up at the bankroll-management level instead.

Does the same Kelly fraction work across domains?

Quarter Kelly is the universal sane default. Some domains (poker, lower-edge horse racing) push toward 0.10-0.20 because of higher variance. Prediction markets with calibrated edges can support 0.50.

Bottom line

The Kelly Criterion is the same formula across horse racing, poker, and prediction markets — but each domain has its own frictions that change how the formula should be applied. Pari-mutuel drift in racing, opponent variance in poker, walk-the-book and resolution lag in prediction markets. Quarter Kelly with robust shrinkage is the universally sane default. A proper Kelly calculator handles the math; the discipline around it is what compounds.

Size Your Polymarket Bets Like a Pro

Try our free Kelly Criterion calculator — Bayesian shrinkage, walk-the-book sizing, 10k-trial Monte Carlo. Built for serious Polymarket traders. Auto-fills from live ProbBrain signals.

🎯 Open the Kelly Calculator →

18+ only. Gamble responsibly.

Search This Blog

Crypto Bonus Hub