A backtest that made money answers the wrong question. The right one is colder: why did it make money — and who was on the other side? Every dollar of excess return is a dollar someone else gave up. They either lost it, or they handed it over willingly in exchange for something they wanted more. If you can't name that counterparty and say why they stay in that losing position, you don't have an edge — you have a coincidence that happened to hold across your test data.
There are exactly four sources of real return. There is no fifth.
1. Risk premium
Why it exists. The return is compensation for absorbing genuine pain — violent volatility, a drawdown that grinds on for months, and a loss that lands at the worst possible moment, exactly when you need the money most.
Who's on the other side. An investor who can't carry that pain: a short horizon, or obligations that won't wait. They pay you to take the risk off their hands.
On the platform. A strategy held at a 25% volatility target with a 29% maximum drawdown — the return is the wage for enduring that, not a reward for clever timing.
Persistence: high. It never disappears, because it isn't a market error — it's a fair price. The catch: it pays zero in the bad years, which is exactly when you're most tempted to abandon it.
2. Behavioral constraint
Why it exists. Human biases are stable across decades. Fear sells the bottoms, greed buys the tops, and the memory of a crisis fades within about two years — so the same mistake is made again, in full.
Who's on the other side. The trader deciding under emotional pressure — which is the overwhelming majority in any market open to the public.
On the platform. 736 fear days against 330 greed days; an average buy at $7,343 against an average sell at $23,217. The gap is discipline, not prediction.
Persistence: high. It weakens as automation spreads, but it never vanishes, because its source is people, not programs.
3. Structural friction
Why it exists. Large players are forced to act sub-optimally by things they don't control — written investment mandates, index weightings, tax rules, industrial inventory cycles.
Who's on the other side. An institution executing a contractual obligation. It isn't hunting for the best price, and it isn't free to choose its own timing.
On the platform. The gold-to-silver ratio stretches to extremes because half of silver's demand is industrial, on a cycle that has nothing to do with monetary pricing.
Persistence: medium. It lasts until the rule that created the friction changes — a new regulation, a new mandate — and then it is gone.
4. Processing advantage
Why it exists. Cleaner data, sharper modelling, or the speed to catch what isn't priced yet. It comes from tooling and engineering discipline — not from secret information nobody else has.
Who's on the other side. Anyone working with older information or a blunter tool: measuring risk with one fixed rule where you measure it with many.
On the platform. Measuring risk across eight dynamic zones instead of a single fixed threshold — a precision advantage in processing, not in privileged information.
Persistence: low. The shortest-lived and most expensive of the four: every imitator erases part of it. You keep it only by staying ahead.
The four at a glance
| Source | Who pays you | Persistence | Ends when |
|---|---|---|---|
| Risk premium | Investors who can't carry the pain | High | Never — but pays zero in bad years |
| Behavioral constraint | Emotional traders — most of the public | High | Never fully — only weakens with automation |
| Structural friction | Institutions bound by obligations | Medium | The rule or regulation changes |
| Processing advantage | Anyone on older data or a blunter tool | Low | Imitators copy it |
Every edge has a half-life
Even a real edge decays. Anomalies lose roughly 26% of their strength out of sample, and about 58% once they are published and everyone crowds in. The statistical bar has risen to match: what once cleared a t-stat of 2.0 now needs to clear t > 3.0 to survive the correction for all the strategies that were quietly tested and discarded first.
The lesson isn't despair — it's placement. The two durable sources, risk premium and behavioral, are people-shaped and slow to erode; the two fragile ones, structural and processing, live on borrowed time. Knowing which one you are standing on tells you how much return is left in it.
The first question
So the first question was never "did the backtest work?" It is "why does this return exist at all — and who is paying it?" Answer that honestly and you already know most of what matters: whether the edge is a fair price you'll be paid for enduring, a human mistake that keeps repeating, a rule someone is trapped by, or a head start that is already being copied. That single answer — not the equity curve — tells you whether the result survives past the end of your test data.
The next pieces in the series are about measuring it: which return figure to trust, what each risk-adjusted metric hides, and the specific ways a backtest lies.