"Position Sizing and Daily Loss Limits: The Mathematics of Survival"

Traders spend the great majority of their attention deciding what to trade and when. The evidence of blown-up accounts — retail and institutional alike — suggests the attention is misallocated: the variable that most reliably separates survivors from casualties is how much. Position sizing is not a detail appended to a strategy; over any horizon long enough to matter, it is the strategy's risk profile. And unlike signals, sizing is governed by arithmetic that is fully knowable in advance. This article walks through that arithmetic and ends where it logically leads: hard-coded loss limits.

The asymmetry everyone knows and few internalize

Losses and gains are not symmetric in their demands. A 10% drawdown requires an 11% gain to recover; 25% requires 33%; 50% requires 100%; 75% requires 300%. The function is convex and unforgiving: each additional unit of drawdown costs disproportionately more future performance to repair. Two consequences follow directly. First, volatility itself is a cost — two strategies with identical average returns but different variance compound differently, and the more volatile one compounds less (the gap, roughly half the variance, is the "volatility drag"). Second, the practical objective of sizing is not to maximize the best case but to keep the drawdown distribution inside the region from which recovery is realistic — mathematically and psychologically, because decision quality degrades precisely as drawdowns deepen.

Kelly: the ceiling, not the target

The theoretical anchor of sizing is the Kelly criterion (Kelly, 1956): the fraction of capital that maximizes the long-run growth rate of wealth for a bet with known edge and odds. Kelly is intellectually indispensable — it proves that optimal size is finite, that it scales with edge and inversely with variance, and that betting beyond Kelly strictly reduces growth while increasing risk: past the optimum, more aggression buys you less money and more pain.

But applying full Kelly to trading collides with three realities. Your edge is estimated, not known — and Kelly is brutally sensitive to overestimated edges, which, per our discussion of backtest overfitting, is the default state of estimated edges. Returns are not independent coin flips — regimes cluster losses. And full-Kelly drawdowns are savage even when the inputs are exactly right (a full-Kelly bettor should expect to visit a 50% drawdown). Hence the near-universal professional practice of fractional Kelly — sizing at a quarter to a half of the computed optimum — which sacrifices modest theoretical growth for a dramatic reduction in drawdown depth and in sensitivity to estimation error. The honest summary: compute Kelly to know where the cliff is; stand well back from it.

From theory to desk practice: volatility-based sizing

The workhorse implementation in futures trading is risk-parity-style volatility targeting at the position level: define the dollar loss you are willing to risk per trade or per day, measure the instrument's current volatility (ATR or realized vol), and size so that a normal adverse move costs your budgeted amount — not more. The discipline this imposes is automatic and countercyclical: positions shrink when markets turn violent and expand when they calm, keeping the risk (rather than the contract count) constant. Its failure mode is equally well known — volatility estimates lag, and regime breaks arrive faster than estimators adapt — which is exactly why sizing rules need a backstop that does not depend on any estimate being right.

The backstop: daily loss limits as architecture

That backstop is the daily loss limit: a pre-committed maximum loss — per strategy and, critically, per account — beyond which the system stops initiating risk for the session, full stop. Its logic is threefold.

Statistically, it truncates the left tail. Whatever the true distribution of your daily P&L, a hard floor at −X converts unknown tail risk into a known, bounded quantity — the single most powerful transformation available in risk management.

Behaviorally, it removes the decision from the moment least suited to making it. The documented tendency of decision-makers to increase risk while losing (the disposition to "make it back") is the mechanism behind most catastrophic trading days, human and algorithmic alike. A limit decided calmly in advance and enforced by code is the only version that reliably survives contact with a losing streak.

Architecturally, the limit must live in the order path — evaluated pre-trade, on the aggregate account, with breaches triggering the escalation ladder we describe in our kill-switch article (halt new risk; reduce; flatten if configured). A loss limit implemented as an end-of-day report is a diary, not a control. This is why GIDEON enforces loss floors and position caps at the middleware layer, where every order from every strategy must pass, rather than trusting each strategy to police itself.

The composed system

Sizing and limits are not alternatives; they are layers. Volatility-based sizing sets the expected risk of each position; fractional-Kelly reasoning caps aggregate aggressiveness relative to estimated edge; the daily loss limit bounds the realized outcome when both estimates fail — and they will, occasionally, together. A trading operation with all three layers can be wrong about markets many times and remain in business. An operation without them needs to be right with a consistency that markets have never granted anyone.

Survival is not the reward for good trading. It is the precondition.

References

• Kelly, J. L. (1956). "A New Interpretation of Information Rate." Bell System Technical Journal, 35(4).
• Thorp, E. (2006). "The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market." In Handbook of Asset and Liability Management. Elsevier.
• MacLean, L., Thorp, E. & Ziemba, W. (2010). "Good and Bad Properties of the Kelly Criterion." Quantitative Finance, 10(7).
• Odean, T. (1998). "Are Investors Reluctant to Realize Their Losses?" Journal of Finance, 53(5).