Why most backtests flatter
A simulated track record answers an easier question than the one you care about. Here is how to read one without being fooled — and why this site labels them the way it does.
A backtest is a simulation of a strategy on historical data. It answers the question: had you run these rules in the past, what would the record look like? That is a legitimate question. The trouble is that it is not the question an allocator cares about, which is what the rules will do next — and the gap between those two questions is where most of the flattery lives.
The first source of flattery is selection. Nobody publishes their seventh-best backtest. By the time a simulated record reaches you, it has usually survived a private tournament among dozens of variants — lookback windows, universes, rebalance dates — and the winner won partly because it fit the sample. Researchers call this overfitting; practitioners call it Tuesday. The effect is mechanical: the more configurations you try, the better the best one looks, whether or not anything real is there.
The second source is cost. Simulated fills are polite. They execute at the close, in size, without moving the market, without borrow constraints, without the Tuesday your data vendor was down. Real fills are not polite. For fast-turnover strategies, realistic costs routinely erase a third or more of simulated returns — which is why every methodology disclosure on this platform must state its cost assumptions per market, and why a backtest that omits them is not listed.
The third source is regime luck. A momentum backtest run over a sample dominated by one long expansion is partly a measurement of the strategy and partly a measurement of the weather. One of the agents on this marketplace — Cascadia Momentum — currently shows a backtested Sharpe ratio above 2.5. Its author has been careful where most retail backtests are careless: point-in-time constituents, next-day executions, stated costs. The number is still labeled hypothetical on every surface it appears, because a sample that favored exactly this style cannot testify about the next one.
How to read one anyway
None of this makes backtests worthless. It makes them evidence of a specific, limited kind — and there are tells that separate careful simulation from marketing.
Look for stated costs, per market, and be suspicious of round zeros. Look for the ugly stretch: a simulated record with no losing quarters is telling you about the search process, not the strategy. Look for capacity honesty — a strategy that made 40% a year in single-name equities usually made it in sizes that would not absorb your capital. And look for the author's willingness to say what breaks the model: an edge whose failure mode is unstated is an edge that has not been examined.
What the label does
On Semantic Securities, every figure derived from simulation carries the word hypothetical — on the card, the row, the profile, the chart itself. The label is not decoration; it is the boundary of the claim. When an agent begins emitting signals through the platform, its record forks: from that date forward, performance is computed from platform-timestamped signals the author can no longer edit. The dashed line becomes solid. That boundary — visible on every chart as a dateline — is the difference between a story about the past and a record of the present, and it is the entire reason the verification tier exists.
The honest summary: a backtest tells you a strategy is plausible. A verified record tells you it is real. The marketplace prices the difference, and so should you.