How every result here is computed
A backtest is only as honest as its assumptions — and most of the ones you see online quietly break them to look good. This page is the whole recipe: the rules my engine can't bend, where the data comes from, and — just as important — what a backtest can't tell you. If any result on this site ever contradicts what's written here, the page is wrong, not this page.
Open method — v2 · since October 28, 2026. Everything labelled “tested to our published open method v2” across the site was produced by exactly these rules; a frozen census edition names the revision it used (what changed, and when).
Rules the engine can't bend
These are enforced in code, on every single run — yours and ours.
A signal for today can only use data that had already closed. The engine acts on the next bar's open, never on a price it couldn't have known at decision time. This is the single most common way online backtests cheat; we make it structurally impossible.
Every fill pays 5 basis points (0.05%) per side — commission plus a slippage estimate — charged on entry and exit, so a round trip costs about 0.10%. It's on by default and not removable; a strategy that only works at zero cost doesn't work. Fills execute at the next open after the signal bar closes, and returns annualize on a 252-trading-day year for stocks and ETFs, and a 365-day year for 24/7 crypto.
A stop-loss, ATR stop or trailing stop fills at its level when the day trades down through it — or at that day's open when the day opens past it, so a stop never fills at a price the day didn't trade, and a gap can cost more than the stop's distance. A take-profit is booked at its level even when the day opens past it, and a day that reaches both your stop and your target counts the stop — both lean cautious. (Mirrored for a short.)
Every result sits next to the total-return of simply buying and holding the same asset over the same window. Beating that dashed line is the entire point — and most strategies don't.
The whole equity curve, the max drawdown, the losing trades, and the periods in cash are all shown. Nothing is trimmed to flatter a result. No cherry-picked windows, no hidden losing years.
We never state or imply that any strategy will make money. Results are hypothetical; the value is the honest test, not a promise. (See the disclaimers at the foot of every result.)
We hold no funds and execute no trades. Alerts are a mechanical heads-up that your own rule fired on the close; you decide and execute on your own broker.
Where the numbers come from
Prices are pulled from public, keyless sources: Yahoo Finance for stocks, ETFs, FX and commodities, and for crypto an aggregate USD reference price and volume across major exchanges (also via Yahoo Finance) — not a single exchange's raw prices. Equity is computed on a total-return basis — split-adjusted prices with dividends credited as cash — so a dividend-paying stock is compared fairly against a strategy that sits in and out of it, and a later dividend can never silently rewrite a result that was already recorded.
Before any backtest runs, a data-quality gate checks the underlying series for staleness, gaps, zero-volume/halt bars and improbable single-bar jumps, and flags them on the result — because a clean-looking curve built on bad data is its own kind of lie.
The asset list is a curated set of famous, liquid names. That is a survivorshipchoice — well-known survivors — and I label it as such rather than pretend it's the whole market. What survivorship bias does to a backtest — shown on this data →
Backtest vs forward-test — and what neither can tell you
It runs a fixed rule over history you can already see — which makes it easy to overfit: tune the knobs until the past looks perfect and the future ignores you. A great backtest is a hypothesis, not a result.
A paper forward-test starts flat today and accrues with zero hindsight — it can only ever act on data it couldn't have known. It is still hypothetical / simulated(no real capital, no market impact), but it can't be curve-fit after the fact, which is why it's the closest thing to real proof we offer.
- • Overfitting. More parameters and more conditions make the past look better and predict the future worse.
- • Regime dependence. A rule that shone in one market (a long bull run, a volatile tape) can quietly stop working when the regime turns.
- • Small samples. A great-looking result on a handful of trades is mostly luck. Treat it that way.
- • Hypothetical ≠ live. Real fills, real spreads, taxes, and your own behaviour under a drawdown are not in any simulation.
A backtest is one stage of five
This page is the HOW — how each number is computed. The five stages are the WHEN and WHY.
Testing a strategy is stage two of preparing to risk real money — not the whole job. The full path is Understand → Test → Stress-test → Forward-test → Decide, and most ideas should be killed — for free — in the first three. Reaching the end doesn't mean an idea works; it means you've stopped fooling yourself about whether it does.
"Beat buy & hold" means exactly one thing
Everywhere we say a strategy "beat" or "lagged", we mean its total return over the window was higher or lower than simply holding the asset over the same window — net of costs. Nothing about win rate, and never a claim about the future. It's why our honest-results page can say, plainly, that most famous strategies don't beat holding — that's the finding, not a flaw. Combinations with too few trades to judge are marked thin and left out of the count.
The Honest Backtest Census: How Often Famous Strategies Beat Buy & Hold
This is the deliberately stocks-&-ETFs cut, frozen and dated so a citation never shifts. The live scoreboard on the honest-results page reads differently — it also includes crypto, which beats buy & hold far more often on this set, so its blended number is higher.
“As of September 25, 2026, 25 famous, named trading strategies were each backtested on 36 liquid US stocks & ETFs — 900 strategy-asset combinations, of which 839 had enough trades to judge (61 were too thin and excluded). Of those judged, about 86% failed to beat simply buying and holding the same asset over the same 5-year window — net of costs and slippage, with no look-ahead. This is a survivorship-biased set of well-known assets, measured over the full period with no out-of-sample split, against each asset's own buy & hold, and is not significance-tested. These are hypothetical, simulated results, not advice.”
Full release — all 25 strategies, drawdowns, and the free CC-BY data →
The census as an open, recurring dataset — all editions and how to cite the series →
Why there are no streaks, badges or levels here
There is no daily streak to protect, no points, no badges, no “you're on a roll” confetti, and nothing that rewards you for coming back. That's on purpose. Those mechanics are built to maximise time on site — they reward engagement, and engagement is exactly the wrong thing to reward in a tool about money. The honest outcome here is often that your idea doesn't work and you close the tab — and a product that nudged you to keep tapping instead would be quietly working against you. The only thing worth optimising for is whether the answer you got was true. The learn pages are written the same way: they teach you how to test and how to rule things out, not how to feel like a winner.
The five stages of preparation are a path to work through, not levels to clear: nothing rewards you for reaching the last one, and a strategy that fails a test is shown failing, never quietly checked off. The only thing that ever “progresses” here is real statistical evidence — how much out-of-sample time a forward test has actually earned.
What changed, and when
Each frozen census edition records the revision it was computed under; a result is only ever compared with one computed under the same rules.
- v2 · from October 28, 2026 — A stop-loss, ATR stop or trailing stop that the market opens through is filled at that day's open, not at the stop level — so a stop can't fill at a price the day never traded. Backtests with stops can show lower results than before this date.
- v1 · from September 25, 2026 — Stops were booked at their level even on a day that opened through them.
Built by A.P., one engineer who builds quantitative trading systems. No name to Google and no résumé to take on faith — the method is open, and the scoreboard shows the whole glass: most strategies fail, mine included, and the rare few that genuinely beat buy & hold. Why I built this →