Position Sizing for Algo Portfolios

Two traders can run the same signals on the same instruments over the same period and end with completely different results. Nothing about the edge differs; only the size of each position does. Sizing is the part of a system that converts a statistical edge into an equity curve, and it is the part most retail systems leave as a constant somewhere near the top of the file.

The three sizing rules worth knowing

There are many sizing schemes in the literature. Three of them cover almost every situation a retail systematic trader will encounter, and they form a natural progression in sophistication.

Notice the pattern down that table: each rule buys better risk normalisation at the cost of one more parameter you have to estimate from noisy data. That trade is not automatically worth taking. A well-implemented fixed fractional rule beats a badly estimated risk-parity model comfortably.

What happens to drawdown as size grows

The relationship between size and return is linear. The relationship between size and drawdown is not, and the relationship between size and risk of ruin is violently non-linear. Doubling your risk per trade does not double your worst drawdown; past a point it converts a survivable drawdown into a terminal one.

This is why the theoretically optimal bet size is a bad practical target. The optimum assumes you know your edge exactly. You do not — you have an estimate from a finite backtest, and that estimate is biased upward by every selection decision you made while building the strategy. Sizing at a fraction of the theoretical optimum costs a little expected growth and buys a large reduction in the chance of a fatal drawdown.

Volatility targeting in practice

Volatility targeting inverts the usual question. Instead of asking how many lots to buy, you ask what annualised volatility you want the position to contribute, then solve for quantity. A position in a quiet instrument gets larger; the same risk budget in a turbulent one gets smaller.

import math  def target_lots(    equity: float,    target_risk_pct: float,   # e.g. 0.10 for 10% annualised    annual_vol: float,        # instrument annualised vol, e.g. 0.18    point_value: float,       # rupees per point, e.g. 75 for Nifty    price: float,    lot_size: int,    max_lots: int,) -> tuple[int, str]:    """Lots such that the position contributes roughly target_risk_pct of vol."""    if annual_vol <= 0:        return 0, "no volatility estimate"     # Rupee volatility of one lot over a year.    lot_notional = price * point_value    lot_risk = lot_notional * annual_vol     budget = equity * target_risk_pct    raw = budget / lot_risk     lots = math.floor(raw)    if lots < 1:        # The smallest tradeable position already exceeds the budget.        return 0, f"minimum 1 lot is {raw:.2f}x the risk budget"    if lots > max_lots:        return max_lots, f"capped at {max_lots} (uncapped {lots})"    return lots, "within budget"

The branch that returns zero is the one worth dwelling on. It encodes an uncomfortable fact: for many Indian F&O contracts, one lot is already a larger risk than a small account should take on a single position. The correct response is to not take the trade, and a sizing function that silently rounds up to one lot instead is the mechanism by which small accounts end up over-leveraged.

Sizing a portfolio rather than a trade

Sizing each position correctly in isolation does not size the portfolio correctly. Five positions each risking 1% are not risking 5% if they are correlated — and in Indian markets, correlation across index and large-cap positions rises sharply exactly when it matters.

The correlation assumption is the fragile one. Estimating it from a calm period and applying it during a stressed one is how diversified books turn out to have been one position all along. Treating correlations as higher than measured, particularly for same-direction index exposure, is the conservative error and the right one to make.

Choosing a rule for your strategy

The right rule depends on how many instruments you trade and whether your strategy has a well-defined stop. Those two questions resolve most of the decision.

Getting it into the system

Sizing rules are simple to write and easy to get subtly wrong, and the errors do not announce themselves — an over-sized book looks like a good one right up until it does not. Testing the sizing function against hand-computed cases is unglamorous work that pays for itself the first time a volatility estimate returns something unexpected.