Portfolio Capacity

Every quant strategy has a size beyond which it stops working. A signal that returns 20% a year on $10m may return 5% on $1b and lose money on $10b, not because the alpha decayed, but because the cost of acquiring the positions ate the edge. Portfolio capacity is the AUM at which marginal trading costs equal marginal alpha, the ceiling on how much money a strategy can profitably run. It is the single most important number separating a backtest from a business, and it is where Market Impact and Transaction Costs meet portfolio construction.

Why capacity is finite

Alpha is roughly scale-invariant per dollar, a good signal predicts returns regardless of your size. But trading costs grow with size. To deploy more capital you must hold larger positions, and to build larger positions you must trade larger quantities, and trading larger quantities moves the price against you. Above some size the extra cost of getting in and out exceeds the extra return, and net alpha turns negative. Capacity is that crossover.

The square-root law of market impact

The empirical regularity, confirmed across markets (Almgren et al. 2005; Kyle's model), is that the price impact of trading a quantity $Q$ scales with the square root of participation:

$\text{impact} \;\approx\; c\,\sigma\,\sqrt{\frac{Q}{V}},$

where $\sigma$ is the asset's volatility, $V$ is its daily traded volume, $Q/V$ is your participation rate, and $c$ is an asset-class constant of order 0.5–1. The cost is concave in size (each extra share is cheaper at the margin than linear would suggest) but still rising, and the total cost you pay, $\text{impact}\times Q \propto \sigma Q^{3/2}/\sqrt{V}$, grows faster than proportionally with position size. Double your AUM and impact cost per trade rises by $\sqrt{2}$; total dollar cost rises by $2^{3/2}\approx 2.83$.

The capacity trade-off

Let a strategy with capital $A$ have gross annual alpha $\alpha_0$ (per dollar) and require annual turnover $\tau$ (dollars traded per dollar of capital, per year). Dollars traded per year are $\tau A$. Spreading this across trades at participation rate $p = (\text{trade size})/V$, the per-dollar impact cost scales as $\kappa\sqrt{A}$ for some constant $\kappa$ absorbing $\sigma$, $V$, turnover and the impact coefficient. Net alpha per dollar is

$\alpha_{\text{net}}(A) = \alpha_0 - \kappa\,\tau\,\sqrt{A}.$

Net dollar profit is $A\,\alpha_{\text{net}}(A) = \alpha_0 A - \kappa\tau A^{3/2}$. Maximizing over $A$ (set the derivative to zero: $\alpha_0 - \tfrac32\kappa\tau A^{1/2}=0$) gives the profit-maximizing size

$\boxed{\;A^\star \propto \left(\frac{\alpha_0}{\tau}\right)^{2},\;}$

and net per-dollar alpha vanishes at $A_{\max} = (\alpha_0/\kappa\tau)^2 = \tfrac94 A^\star$. Two structural facts fall out. Capacity scales with the square of gross alpha, a strategy with twice the edge holds four times the capital, and it scales with the inverse square of turnover: fast strategies capacity-cap first. A signal that turns over 50 times a year has $\sim 100\times$ less capacity than the same-alpha signal turning over 5 times a year.

Worked example

A stat-arb strategy runs $500m at 30% gross alpha with turnover $\tau = 20\times$/year (trades $10b annually). Suppose at this size impact costs run 10% of capital a year, so net alpha is $30\% - 10\% = 20\%$, healthy. Now scale to $2b (4×). Impact cost per dollar scales as $\sqrt{A}$, so it doubles to $\sqrt{4}\times 10\% = 20\%$; net alpha falls to $30\% - 20\% = 10\%$. Push to $4.5b (9×): cost scales by $\sqrt{9}=3$ to $30\%$, and net alpha hits zero, that is $A_{\max}$. Net dollars peaked earlier, at A^\star = \tfrac49 A_{\max} = \2$b, where net alpha is$\alpha_0/3 = 10% on \2b $=$ $200m, versus $100m net at the tiny $500m size. Beyond $2b you manage more money for less profit; beyond $4.5b you lose money. The manager's rational cap is around $2b even though the strategy still "makes money" up to $4.5b.

Levers that expand capacity

• Lower turnover. Since capacity $\propto \tau^{-2}$, halving turnover roughly quadruples capacity, the highest-leverage change. Longer holding periods, wider rebalance bands, and trade-netting all help.
• Trade slower / smarter. Execute over more days at lower participation (impact is convex in speed), use scheduling algorithms (VWAP/IS), cross in dark pools, reducing $\kappa$ directly.
• Broaden the universe. Spreading capital across more, more-liquid names raises aggregate $V$; capacity is additive across independent liquid bets (the breadth term in the fundamental law).
• Optimize net of costs. Build the portfolio with a transaction-cost penalty in the objective (see Pitfalls of Mean-Variance Optimization on cost-aware optimization), trading only when expected alpha exceeds expected impact.

Failure modes

• Backtests ignore impact. A frictionless backtest reports gross alpha and infinite capacity; the strategy dies on contact with real markets. Always simulate net of a realistic $\sqrt{\cdot}$ impact model.
• Crowding. Capacity is shared: if many funds run the same signal, aggregate participation is what moves prices, so your effective capacity is far below your standalone estimate, and unwinds become correlated (the 2007 quant quake, Alpha Decay).
• Liquidity is state-dependent. $V$ collapses in stress exactly when you need to trade; capacity computed on calm-market volumes overstates true capacity.
• Nonlinear past a point. The square-root law is a fit to modest participation; at very high participation impact becomes worse than square-root and can be effectively unbounded.

In interviews

Frame capacity as the size where marginal trading cost equals marginal alpha, alpha per dollar is scale-invariant but impact grows with size, so net alpha eventually crosses zero. Cite the square-root impact law (impact $\propto \sigma\sqrt{Q/V}$) as the empirical backbone, and derive that net dollar profit $\alpha_0 A - \kappa\tau A^{3/2}$ is maximized at a finite $A^\star$. The two punchlines to remember: capacity scales like the square of gross alpha and the inverse square of turnover, so high-turnover strategies capacity-cap first, which is exactly why high-frequency edges are small-AUM and slow value strategies scale to hundreds of billions. See Transaction Costs and Market Impact.