Estimating the Capacity of an Alpha

Every alpha signal has a capacity: an amount of capital above which trading it moves prices enough to eat the return it's chasing. A signal is not a fixed number — "this strategy earns 12% a year" is only true at a specific size. Below capacity it earns 12%; well above it, the same strategy might earn 2% or lose money, because your own trades are now the market-impact cost that erodes the edge. Estimating capacity is the step that turns "does this work" into "how much can we actually run."

Where the capacity limit comes from

The mechanism is simple: trading a signal means buying names it ranks well and selling names it ranks poorly, and that buying and selling pushes prices — a cost that scales with how much of a stock's daily volume you need to trade (participation rate). A larger book needs to trade a larger dollar amount in the same names to keep the same relative positions, which raises the average participation rate, which raises the average impact cost. The signal's gross return is roughly constant with size; the cost is not. Capacity is the AUM at which the rising cost curve meets the (roughly flat) gross alpha curve.

$\text{net alpha}(\text{AUM}) \approx \alpha_{\text{gross}} - c \cdot \left(\frac{\text{AUM}}{\text{ADV}}\right)^{p}$

In words: net alpha starts near the gross alpha estimated in research, and falls off as assets under management grow relative to average daily volume (ADV), at a rate governed by a cost coefficient $c$ and an exponent $p$ typically a bit above 1 — impact costs rise faster than linearly as you demand a bigger share of a stock's daily trading.

Estimating it in practice

Three complementary approaches, used together rather than any one alone:

1. Trade-size participation limits. Cap each name's daily trade at some fraction of its ADV (commonly 5–10% for a strategy that needs to turn over regularly) and see what AUM that cap implies given the signal's position sizes and turnover. This is a hard, defensible ceiling, but it says nothing about whether returns hold up below that ceiling.

2. A market-impact model. Plug the strategy's implied trade sizes into a standard impact model (square-root-in-volume models are common) at several candidate AUM levels, and subtract the modelled cost from the gross alpha at each level. This produces the net-alpha-versus-AUM curve directly, and it is where the capacity number usually comes from in a formal writeup.

3. Live evidence from correlated books. If a similar signal is already traded elsewhere in the firm, its realised slippage at known trade sizes is the best calibration available for the impact model's parameters — model-implied costs are only as good as the coefficients fed into them.

A worked example

A signal shows a gross Sharpe of 1.8 in research, trading names averaging $40 million ADV, with the strategy needing to trade roughly 15% of the position size in each name daily to maintain target weights. Using a square-root impact model calibrated from the desk's own historical fills ($c \approx 0.6$, in basis points, for this cost model, with volatility and volume as inputs), the modelled net Sharpe by AUM comes out roughly as:

AUM	Avg. participation rate	Modelled net Sharpe
$50 million	0.5% of ADV	1.7
$300 million	3% of ADV	1.4
$1 billion	10% of ADV	0.7
$2.5 billion	25% of ADV	0.1

If the desk's hurdle is a net Sharpe of at least 1.0, this signal's estimated capacity is somewhere between $300 million and $1 billion — not the headline $1.8 Sharpe the raw research reported, and not a single hard number either, but a range with a stated hurdle attached.