From Blackjack Tables to Wall Street: Thorp on Actually Using the Kelly Criterion

Finding a bet with a positive edge is hard. Most people assume that once you have found one, the hard part is over. It is not. The second problem, how much to bet, will destroy you just as thoroughly as having no edge at all.

Bet too little and you leave nearly all of the growth on the table. Bet too much and you go broke with probability one, even with a genuine edge, because a large enough losing streak wipes out a base you can never rebuild.

Ed Thorp is the person who took this problem seriously, from both ends. He used the Kelly criterion to beat blackjack, then to beat a casino sports betting operation, then to run a successful quantitative hedge fund for decades. His survey paper is the practitioner's account: what Kelly promises, what it costs, and what he actually did.

The problem: an edge is not a strategy

Say you have a coin that lands heads 55 percent of the time, and you can bet on heads at even money. A clear, genuine edge. Now, what fraction of your bankroll do you bet each flip?

Bet everything. You will go bust the first time it lands tails, which is 45 percent likely on flip one, and essentially certain if you keep doing it.

Bet a tiny amount, say 0.1 percent. You will never go broke, and your money will grow so slowly that you will die before it matters.

Somewhere between those extremes is a fraction that grows your wealth as fast as it can be grown. The Kelly criterion, from John Kelly's 1956 information theory paper, names that fraction: bet the fraction that maximizes the expected logarithm of your wealth. For the simple coin case, that comes out to your edge divided by the odds, which for the 55 percent coin at even money is 10 percent of your bankroll.

The key idea via analogy: compounding is multiplicative, and that changes everything

Here is the intuition that makes Kelly click, and it is the single most useful thing in this whole area.

When you reinvest your winnings, your outcomes multiply, they do not add. A 50 percent loss followed by a 50 percent gain does not leave you flat. It leaves you down 25 percent. Losses hurt more than equivalent gains help, not because of psychology, but because of arithmetic.

This means the thing you should be maximizing is not your expected wealth. Maximizing expected wealth tells you to bet everything on every favorable bet, which is suicidal. The thing to maximize is your expected growth rate, and because growth compounds multiplicatively, that means maximizing the expected log of wealth. Logarithms turn multiplication into addition, which is exactly the transformation the problem needs.

Thorp's paper lays out what this criterion buys you:

• It maximizes the long-run growth rate of your wealth. Over enough repetitions, a Kelly bettor will end up with more money than someone using any essentially different strategy. That is not a heuristic claim, it is a theorem.
• It never risks ruin. Because you always bet a fraction of your bankroll, you can shrink but you cannot hit zero. A Kelly bettor who loses ten in a row simply has a smaller stake and keeps playing.
• It gets you to a target fastest. Kelly also asymptotically minimizes the expected time to reach a given wealth goal.

Those are genuinely strong optimality properties. And Thorp is the proof of concept: he applied it at the blackjack table, in a sports betting system, and in the markets.

The part practitioners actually need to hear

Thorp is refreshingly blunt about the cost, and it is the reason so many people who learn Kelly then misuse it.

Full Kelly is brutally volatile. The growth-optimal fraction is far more aggressive than nearly anyone's stomach or career can survive. A full Kelly bettor can expect deep, sustained drawdowns as a matter of routine. There is a well-known result that a full Kelly strategy will, at some point, halve your bankroll with substantial probability. Sit with that. Not a small chance. A large one.

So Thorp advocates what practitioners overwhelmingly do: fractional Kelly, typically half Kelly. Bet half the growth-optimal amount.

The trade is remarkably favorable, and this is the number worth remembering. Because the growth rate curve is flat near its peak, cutting your bet size to half of Kelly costs you only about a quarter of your growth rate, while cutting your volatility roughly in half. You give up 25 percent of the theoretical growth to remove 50 percent of the pain. Almost everyone should take that trade.

And there is a deeper reason to bet less than Kelly says. Kelly assumes you know your edge exactly. You do not. Your edge is estimated from limited data and may be smaller than you think, or zero, or negative. Overestimating your edge means overbetting, and overbetting is the failure mode that kills. Betting less than Kelly is the honest response to the fact that your edge is itself uncertain, and Thorp is clear that in markets, unlike in blackjack, you never really know your edge.

That gap between blackjack and markets is the through-line of the paper. In blackjack, the deck is a known object and the edge can be computed. In markets, the "deck" is unknown, changing, and full of other people trying to work out the same thing.

Why it mattered

• It made Kelly usable rather than merely elegant. Kelly's original paper is an information theory result about a hypothetical wire and a hypothetical bettor. Thorp is the bridge to the real world: fractional Kelly, edge uncertainty, correlated bets, multiple simultaneous positions, and what happens when you have to estimate everything.
• It gave position sizing a principled foundation. Most traders size positions by feel or by rules of thumb. Kelly, and Thorp's practical version of it, gives you a defensible answer with a clear logic: your bet size should scale with your edge and shrink with your uncertainty.
• It is the same formula as everything else. Edge divided by variance. It is Merton's share, it is the Treynor-Black weight, it is the shape of the maximum-Sharpe portfolio. Thorp's contribution is showing that the same principle governs a card table and a hedge fund.
• The track record is the argument. Thorp did it. He beat blackjack, beat sports betting, and ran a fund with a remarkable record. There is no substitute for a theorist who staked his own money on the theory.

The honest limitations

• The edge is the whole game and you do not know it. Every optimality property of Kelly is conditional on knowing your true probabilities. In markets you are estimating them, and a modest overestimate of your edge translates into serious overbetting.
• The volatility is real, and careers are short. The theorems are about the long run. A professional manager who takes a full Kelly drawdown will likely be fired before the long run arrives. Thorp's own solution, betting a fraction of Kelly, is an admission that the mathematically optimal strategy is not the practically optimal one for a human being with clients.
• Correlation across bets is not optional. Kelly sizing applied independently to twenty correlated positions is enormous, hidden overbetting. When your bets move together, they are effectively one big bet, and you must size accordingly.
• It assumes you can keep playing. Kelly's no-ruin guarantee relies on infinitely divisible bets and no minimum bet size, no margin calls, and no investors who redeem after a bad quarter. Real constraints can turn a survivable drawdown into an actual ruin.
• Log utility is a choice, not a law. Kelly is optimal for someone who maximizes log wealth. Not everyone should. An investor with a fixed liability, a short horizon, or a lower tolerance for pain rationally bets less, and there is nothing wrong with that.

The one-line takeaway

Thorp took Kelly's growth-optimal betting formula out of information theory and into casinos and markets, and his practical verdict is the one to remember: size your bets in proportion to your edge and inversely to your risk, then bet noticeably less than the formula says, because the formula assumes a certainty about your edge that you will never actually have.