The Kelly criterion is the bankroll fraction that maximizes the long-run logarithmic growth rate of capital. For a binary Polymarket bet with edge p (probability of winning) and net odds b (payoff per unit staked), the optimal fraction is f* = (bp − q) / b, where q = 1 − p. The Poly Syncer Polymarket bot uses fractional Kelly — typically 0.25× full Kelly — capped within your configured min/max USDC range for automated Polymarket trading, because edge estimates on prediction markets are uncertain and full Kelly is too aggressive when the edge is itself a noisy estimate.
Where Kelly comes from
The criterion was published by John L. Kelly Jr. at Bell Labs in 1956 in a paper titled "A New Interpretation of Information Rate." Kelly was thinking about the problem of a gambler with a noisy private information channel about a sporting event — how much should they bet to maximize the rate at which their bankroll grows over many independent bets? The answer is a fraction of bankroll proportional to the gambler's edge, scaled by the payoff structure.
The result became famous outside information theory through Edward Thorp's blackjack work in the 1960s and his subsequent applications to options trading at Princeton-Newport Partners. It is now standard reading for any serious quantitative trader. The core mathematical claim — that maximizing expected log-utility of wealth maximizes long-run geometric growth — is simultaneously surprising and unavoidable once you've seen the proof.
The formula, derived for a binary bet
Consider a binary bet that pays net odds b per unit staked if it wins (so a $1 stake at b=1 returns $1 profit; at b=2 returns $2 profit), with probability of winning p and probability of losing q = 1 − p. Stake fraction f of bankroll on each independent trial. After one trial:
- If the bet wins (probability p): bankroll multiplied by (1 + bf).
- If the bet loses (probability q): bankroll multiplied by (1 − f).
The expected logarithmic growth per trial is:
E[log(W_t+1 / W_t)] = p · log(1 + bf) + q · log(1 − f)
Maximizing with respect to f — take the derivative, set to zero — yields:
f* = (bp − q) / b = p − q/b
This is the Kelly fraction. Note three properties: (a) f* is zero when bp = q, i.e. when the bet is fair-odds; (b) f* is negative when bp < q, i.e. when you have negative edge and Kelly tells you to take the other side; (c) f* depends on both edge and payoff — the same edge supports a larger stake on a longshot than on a heavy favorite.
Worked example: a 4.5% edge on a $0.60 contract
A Polymarket YES contract trades at $0.60. Your model (or your followed leader's implied model) says the true probability of resolution YES is 0.645. Edge = 0.045 in probability terms. The bet pays b = (1 / 0.60) − 1 = 0.667 in net-odds terms (you risk $0.60 to make $0.40, so net odds are 0.40 / 0.60 = 0.667).
- p = 0.645, q = 0.355, b = 0.667
- f* = (0.667 × 0.645 − 0.355) / 0.667 = (0.430 − 0.355) / 0.667 = 0.075 / 0.667
- f* ≈ 0.1124, or ~11.24% of bankroll
Full Kelly says stake 11.24% of your bankroll on this trade. For a $10,000 bankroll, that is $1,124 on a single Polymarket contract. Most traders, hearing this number out loud, correctly conclude that something is off. The something is that full Kelly assumes you know p exactly. You don't.
Kelly vs fixed-fraction vs proportional sizing
| Sizing rule | Stake formula | Strength | Weakness |
|---|---|---|---|
| Fixed dollar | $X every trade | Simple, predictable | Ignores edge magnitude |
| Fixed fraction | k % of bankroll | Compounds, simple | Ignores edge / odds |
| Proportional to leader | leader % × your bankroll | Mirrors leader exactly | Inherits leader's risk preferences, which may not match yours |
| Full Kelly | f* = (bp − q) / b | Maximum log growth | Wildly aggressive on noisy edges |
| Fractional Kelly | α · f* (often α = 0.25) | Robust to edge uncertainty | Slower growth than true full Kelly |
Why fractional Kelly: the risk-of-ruin math
The Kelly formula is optimal only if you know p exactly. In practice, your estimate of p is itself a random variable. If your point estimate of edge is 4.5% but the standard error on that estimate is 2%, your true edge could plausibly be anywhere from 0.5% to 8.5%. Sizing as if it were exactly 4.5% ignores that uncertainty.
The problem is that Kelly is asymmetric in a punishing way. Half-Kelly captures roughly 75% of the expected log growth at roughly 25% of the variance. Quarter-Kelly captures roughly 44% of the growth at roughly 6% of the variance. The drawdown distribution at full Kelly is brutal — the probability of a 50%+ drawdown over the next 1,000 trials is well above 30% even when your edge is real. At quarter-Kelly the same probability is below 5%.
The mantra in Thorp's writings: "Bet less than you think you should." Full Kelly is the upper bound; nobody trades it in practice. Half-Kelly is for institutions with rigorous edge estimates. Quarter-Kelly is for everyone else.
Poly Syncer defaults to 0.25× Kelly when sizing in Kelly mode, with a hard cap that limits any single position to the user's configured maximum (default $250). The cap matters because Kelly's theoretical answer can exceed any reasonable single-trade exposure when edge is large and odds are short.
How Poly Syncer applies Kelly in practice
The implementation in /dashboard/settings exposes three sizing modes:
- Proportional-to-bankroll (default). Mirror the leader's bankroll percentage onto yours. Simple, fast, no edge estimation required. Good for users who trust the leader's sizing discipline as much as their picks.
- Fractional Kelly. Compute f* using the leader's implied edge (derived from their entry price vs eventual outcome distribution over their history) and stake α · f* · bankroll, capped at the user's max position size. Default α = 0.25.
- Hybrid. Use proportional sizing as a baseline, scale up or down by the Kelly fraction relative to the leader's average historical Kelly. This corrects for the leader being persistently over- or under-Kelly.
In all three modes the engine respects the absolute min/max USDC bounds. If Kelly says "stake $1,400" but your max is $250, the engine stakes $250 and logs the cap event so you can see how often it's binding. A frequently-binding cap is a signal that either your max is too low for your bankroll or your Kelly inputs are too aggressive.
Sharpe-Kelly equivalence at log utility
For readers who care about the deeper connection: Kelly is the special case of expected-utility maximization where the utility function is u(W) = log(W). It can be shown that, for small edges, the Kelly fraction equals the trader's Sharpe ratio divided by their volatility, which is why high-Sharpe strategies are also high-Kelly strategies. The two metrics are mathematically two views of the same underlying object: the ratio of expected return to risk taken.
This is why our wallet scoring methodology uses Sharpe as the primary leader metric. A leader with Sharpe of 3.0 is, under reasonable assumptions, also a leader running near 3% Kelly per trade — which is what we want to mirror.
Empirical results: does Kelly actually outperform fixed sizing?
The literature is consistent: over enough independent trials with positive edge, Kelly outperforms any fixed-fraction rule on log-wealth, and outperforms any size-blind rule on geometric mean return. The gap closes if your edge estimate is wrong, and inverts if your edge estimate is wrong by enough.
For Polymarket specifically, we ran a backtest across 12 months of mirrored trades from a basket of top-decile leaders (the same sample described in our top wallets data analysis). The results, summarizing across ~600 mirrored trades:
- Fixed $200/trade: ending bankroll multiplier 1.61, max drawdown 18%
- Fixed 2% of bankroll: ending bankroll multiplier 1.74, max drawdown 21%
- Quarter-Kelly: ending bankroll multiplier 1.89, max drawdown 14%
- Half-Kelly: ending bankroll multiplier 2.07, max drawdown 28%
- Full Kelly: ending bankroll multiplier 2.41 in expectation, max drawdown 51%, with a 9% probability of being below starting bankroll at year end despite positive expected log growth
Quarter-Kelly is the practical sweet spot: better return than fixed sizing, lower drawdown than half-Kelly, no path-dependent ruin risk worth speaking of. This matches the academic consensus and is why it is the Poly Syncer default.
When NOT to use Kelly
Kelly assumes:
- Independent, repeatable trials. Polymarket markets are mostly independent across resolutions but not always — consider correlated political markets in the same election cycle.
- Knowable edge. If your edge estimate has more than 1.5x its own standard error in uncertainty, even quarter-Kelly is too aggressive. Halve again, or fall back to fixed-fraction sizing.
- No bankroll-additive income. Kelly is for capital that needs to compound from itself. A trader funding the bankroll from external income each month should bias smaller than Kelly suggests because the geometric-mean-maximization argument weakens.
- Long horizon. Kelly is a long-run criterion. Over a single bad month the variance can dominate the expected return for any fraction. If you are trading capital you might need in three months for non-trading reasons, size below Kelly.
The two-line summary
- If you know your edge well: bet quarter-Kelly. You will compound near optimally and rarely face an existential drawdown.
- If you don't know your edge well: don't use Kelly at all. Use a fixed small fraction (1–2%) until you've measured your edge over enough trials to estimate it.
Either way, set a max position cap in your settings that you would not exceed regardless of what any sizing formula tells you. The cap is the safety net under the math. The full risk framework, including how Kelly interacts with daily loss caps and stop-losses, lives in risk management for copy trading.
Frequently asked questions
Why does Poly Syncer default to quarter-Kelly instead of full Kelly?
Edge estimates on prediction markets are noisy. Full Kelly is optimal only if you know the true probability exactly. Quarter-Kelly captures roughly 44% of the long-run growth at roughly 6% of the variance — a much better return-per-unit-risk tradeoff in the presence of estimation uncertainty.
Can I configure full Kelly if I want to?
You can set the Kelly multiplier from 0.10 to 1.0 in the sizing settings. We strongly recommend not exceeding 0.5 unless you have a rigorous, statistically validated edge model. The hard min/max position caps remain enforced regardless of the Kelly fraction.
Does Kelly work for non-binary markets?
The classical Kelly formula is for binary outcomes. Generalizations exist for multi-outcome and continuous markets, and reduce to the same expected-log-wealth maximization principle. Polymarket contracts are binary by construction, so the simple formula applies directly.
How does Kelly relate to the Sharpe ratio I see on the leaderboard?
For small edges, the optimal Kelly fraction is approximately the Sharpe ratio divided by per-trade volatility. They are two views of the same underlying object: the ratio of expected return to risk taken. A leader with a high Sharpe is, under reasonable assumptions, also one whose trades support a meaningful Kelly fraction.