Risk Models Ch. 2: Sizing in the Real World — Kelly Without Clean Odds

Investment Background

Chapter 1 established the formula and its central lesson: overbetting a real edge can still cause ruin. That chapter used a clean example — a 55% coin flip with even-money payoffs — because clean numbers make the math legible. Real markets never hand you numbers that clean.

This chapter covers the gap between the textbook formula and using it (or, more precisely, using its logic) on an actual portfolio: how to estimate win rate and payoff ratio from historical returns, what happens when your bets are not independent of each other, and the specific ways the formula's assumptions break down outside the casino.

The Wall Street Translation

Estimating the Inputs When There Is No Dealer

A blackjack card-counter knows the exact odds at every count — the deck is a closed, fully specified system. A trading strategy or a factor tilt is not. You must estimate W (win probability) and R (payoff ratio) from historical data, and every historical sample is finite and non-stationary — meaning the true odds may have already shifted since your data was collected.

The continuous-outcome version of the Kelly formula (used when returns are not binary win/lose but a distribution of outcomes) requires the mean and variance of that return distribution:

f* = μ / σ²

Where μ is the expected excess return of the strategy and σ² is its variance. This version is more relevant to portfolio-level decisions than the coin-flip formula, because most real exposures — a factor tilt, an asset class, a trading strategy's return stream — produce a range of outcomes, not a binary win or lose.

The practical problem: μ and σ² estimated from 5 or 10 years of data carry enormous uncertainty, especially μ. Mean returns are notoriously hard to estimate precisely even from long histories, while variance is comparatively easier to estimate well. This asymmetry is itself an argument for fractional Kelly — you can trust the denominator of the formula much more than the numerator, which means your f* estimate inherits most of its error from the part you are least sure about.

Correlated Bets: Why Your "Edges" Are Not Independent

The simple Kelly formula assumes each bet is independent of the others — one coin flip's outcome does not affect the next. A real portfolio violates this constantly.

Suppose you run three separate strategies, each individually Kelly-sized at what looks like a safe fraction. If all three strategies tend to lose money in the same conditions — a market-wide liquidity crunch, for instance — your effective combined bet size is much larger than any single strategy's fraction suggests, because the losses arrive together rather than diversifying each other away.

Scenario	What single-strategy Kelly sizing misses
Three "independent" equity strategies, all long-only	All three lose simultaneously in a broad market drawdown — effective correlation is high even if the strategies' logic differs
A factor tilt plus a leveraged ETF plus margin debt	All three are forms of the same underlying bet (increased market exposure) wearing different costumes
Genuinely diversifying strategies (e.g., a trend-following overlay that tends to gain during equity crashes)	Correctly reduces effective portfolio-level risk — this is the one case where combining bets can let you safely size larger than any single-strategy calculation implies

The rule this produces: compute Kelly sizing at the portfolio level, using the correlation structure of your actual combined exposures — not by Kelly-sizing each position in isolation and summing. Summing isolated Kelly fractions systematically overstates how much you can safely bet whenever your bets share a common risk factor, which for most retail portfolios, most of the time, they do.

Where the Formula Quietly Breaks

Three assumptions Kelly depends on, and what happens when markets violate them:

1. Stationarity — the odds don't change bet to bet. Markets have regimes; a strategy's edge in one regime can vanish or reverse in the next. A Kelly fraction computed on stale data can be sizing for an edge that no longer exists.

2. Known, fixed payoff structure — Kelly assumes you know the full distribution of outcomes in advance. Markets produce fat-tailed, occasionally catastrophic outcomes that historical data systematically underrepresents (this is Against the Gods' and Antifragile's territory — measurable risk versus unmeasurable uncertainty). A formula built on a known distribution is only as good as your estimate of that distribution, and tail events are precisely what your estimate is worst at.

3. The ability to actually take the position size the formula recommends — real markets have liquidity limits, margin requirements, and taxes that a clean formula ignores.

None of this makes Kelly useless. It makes Kelly a discipline for thinking about sizing, not a black box you feed numbers into and trust blindly — which is exactly why fractional Kelly, and a wide margin of humility about your inputs, is standard practice rather than a hedge for the timid.

Executable Trading Rules

1. If you compute a Kelly fraction for any strategy or tilt, compute it from the longest, most representative data you can get — and still discount the result heavily. Treat a "10% Kelly" result as license for perhaps 2-3%, not 5%.

2. Before combining multiple positions or strategies, ask whether they share a common failure mode. If a single bad market event would hurt all of them at once, size the group as one bet, not several independent ones.

3. Re-examine your inputs periodically rather than setting a position size once and leaving it. An edge estimated from 2015-2020 data may not reflect 2026 conditions; stale odds produce stale, and potentially dangerous, sizing.

4. Treat any Kelly-derived number as an upper bound under best-case assumptions, not a target to reach. The formula's own math shows the cost of erring small is minor; the cost of erring large is catastrophic. Asymmetric consequences justify asymmetric caution.

Relevance to a Retirement Portfolio

This chapter's most important lesson for a retirement investor has nothing to do with computing a formula — it is the correlation point.

Most retail investors who believe they are "diversified" are Kelly-sizing several bets that are secretly the same bet. A tech-heavy stock portfolio, a leveraged Nasdaq ETF, and a margin loan used to buy more of the same holdings are three wrappers around one concentrated exposure. Each one, evaluated in isolation, might look like a modest, safely-sized position. Combined, they are a single oversized bet on one outcome — exactly the overbetting scenario Chapter 1 showed can produce ruin even with a real edge.

What looks diversified	What it actually is, in Kelly terms
S&P 500 index fund + a "growth" tech fund + individual FAANG shares	Overlapping exposure to the same handful of large-cap growth names, sized as if they were three separate bets
Home equity + REIT fund + real estate stocks	One real estate/interest-rate bet in three containers
Global stock index + leveraged version of the same index "for extra growth"	The leverage does not diversify anything — it increases the size of the identical underlying bet

The practical takeaway for your retirement account is not "go compute Kelly fractions for your holdings." It is: before adding any new position, ask what it is correlated with in your existing portfolio, and size the combination, not the piece. This is precisely why a genuinely diversified, low-cost core (spanning asset classes and geographies rather than concentrated in one theme) remains the foundation, with any tactical or satellite position sized small enough that its correlation risk to the rest of the portfolio cannot cause the kind of ruin Chapter 1 described.

Kelly does not tell you to avoid risk. It tells you, with mathematical force, that the size and correlation structure of your bets — not their existence — is what determines whether you compound wealth or eventually lose it.

Chapter 3 turns from sizing a single bet to sizing an entire portfolio at once: Risk Parity, and why balancing risk — not capital — across asset classes changes how a portfolio behaves in a crisis.