A Taxonomy of Alpha Sources

Classifying an edge is not an academic exercise. Where a signal sits in the taxonomy is the best predictor of the three things you actually need to forecast: how long it keeps working, how much capital it absorbs, and what evidence ought to convince you it is real. Two signals with identical backtests but different sources should be sized, monitored and defended differently.

The organising question is blunt: who is on the other side, and why do they keep coming back? Every dollar of active return is a dollar somebody else did not make. Before costs, across all participants,

$\sum_i w_i \, \alpha_i = 0,$

where $w_i$ is participant $i$'s share of the market and $\alpha_i$ their return in excess of it. In plain English: active management is zero-sum before fees and negative-sum after. So if your backtest works, someone is systematically on the losing end — and unless you can name them and say why they tolerate it, the most likely loser is your own backtest. There are four durable answers.

The four families

Risk premia. Your counterparty is not losing; they are buying insurance. You hold something with genuinely unpleasant properties — losses that arrive in bad times, when everyone needs cash — and get paid for it. The equity premium, credit spreads, the The Variance Risk Premium and much of carry live here. The payer keeps paying because avoiding the risk is worth it to them, so publication does not kill it. It will hurt exactly when you can least afford it, which is the whole reason it exists.

Structural. Your counterparty is compelled to trade by a rule, mandate or deadline, and profit is not their objective. Index funds must buy an addition on the effective date; a pension must rebalance to policy weights; a dealer must hedge inventory. Nobody is irrational — they are constrained. Persistence depends on the constraint staying in place, and capacity is bounded by the size of the forced flow.

Behavioural. Your counterparty is making a systematic judgement error: underreacting to news, extrapolating trends, overpaying for lottery-like payoffs. Everyone likes this family because the story is fun, and it disappoints most, because an error is only exploitable if arbitrage is limited. Published behavioural edges decay, though rarely to zero, because the Limits to Arbitrage that let the error survive also constrain whoever arbitrages it.

Informational and analytical. Your counterparty does not know yet, or has not done the work. You bought a satellite feed, parsed the filings, or built a better model of the same public data. Nothing structural protects this — it lasts exactly as long as your data or compute advantage, which with a commercial vendor means until enough other funds subscribe. Highest decay rate of the four, usually the lowest capacity.

Family	Who pays	Survives publication?	Typical capacity	Evidence that convinces
Risk premium	Insurance buyers	Yes	Very large	Losses cluster in bad times; stable across decades
Structural	Constrained institutions	Yes, while the rule holds	Bounded by the flow	The flow is measurable and dated in advance
Behavioural	Erring investors	Partly — decays	Moderate	An arbitrage limit explains why it persists
Informational	The uninformed	No	Small to moderate	Others lack the data, and you can say why

Worked example: an index addition

A mid-cap name is added to a major index. Funds tracking it hold roughly 12% of the float of a typical member and must own that stake by the effective close. The stock trades about 1% of shares outstanding on a normal day, so the required buying is twelve days of volume compressed into one.

Use the square-root rule of thumb: impact runs at daily volatility times the square root of order size over daily volume. With 2% daily vol and a size ratio of 12, that is roughly $2\% \times \sqrt{12} \approx 7\%$ of temporary pressure. In the 1990s the measured announcement-to-effective abnormal return was around 3–5% — the right order of magnitude.

Now the taxonomy earns its keep. This is structural, so it should persist as long as the mandate does — and it did, for two decades. It should also shrink as arbitrage capital arrives and index funds spread execution over days rather than crossing at the close. Both happened: the measured index effect is statistically indistinguishable from zero in recent samples. The family told you the edge was real and told you the mechanism of its decay. See Index Rebalance Arbitrage.

Worked example: classifying a mystery signal

A researcher reports that firms whose 10-K risk-factor section changed materially underperform next quarter, net Sharpe 0.9. Which family?

Not a risk premium — losses do not cluster in bad times, and nobody demands compensation for holding firms with new risk language. Not structural — nobody is forced to trade on filing text. Behavioural is possible: inattention to a long, dull document. But the sharper answer is informational-analytical: the data is public and free, and the edge is the cost of processing it. That tells you what to worry about, and it is not the backtest. It is that three vendors now sell a packaged 10-K change score and every client is running the same screen. Your next test is not another robustness check — it is finding out how many people already have this.