There's a chart circulating. Bitcoin divided by the dollar index, on two-week candles, with an ATR band trend system on top. The results panel in the corner (as of July 2026):

10 buys. 10 sells. Win rate 100.00 percent. Return: 34,130,419 percent.

This is either a textbook case of backtest overfitting — or it isn't. That's the question, and it's answerable. So we're rebuilding the system and testing it.

We won't say who posted it. It doesn't matter — this pattern exists a thousand times over, and the person is interchangeable. The pattern isn't.

We could point out that a win rate of exactly 100 percent is suspicious, and we'd be right. But being right is cheap.

So we're doing it the other way around: the rules are in this article. The computation comes after.

What overfitting actually is

Overfitting doesn't mean somebody cheated. It usually happens without intent.

You try parameters. You keep the combination that looks best. You show it. The result no longer measures the market — it measures the search.

The second major error is look-ahead bias: the signal uses data it couldn't have known at the moment of decision. That, too, is usually accidental, and it produces win rates nobody else can reach.

Both errors look identical in the results panel. You don't tell them apart by looking. You tell them apart by computing.

Why the order is the entire argument

If you compute first and choose your standards afterwards, you always find what you were looking for. Not out of malice — but because a standard chosen after seeing the results isn't a standard anymore. It's a story.

Research calls the countermeasure pre-registration: you state publicly what counts as passing before you look at the data. You bind yourself.

In trading, this essentially never happens. Every screenshot shows a configuration that worked. Nobody shows the thirty that didn't.

So: here are the rules. They don't get touched afterwards.

What we're rebuilding — and what we can't

We don't know which indicator runs in the original. A screenshot won't tell you. We're building an ATR band system consistent with what the chart appears to show, and we're calling it what it is: a reconstruction, not the strategy.

That isn't a weakness of our test. It's the first finding. A result whose parameters nobody knows is a result nobody can verify — including the person showing it.

Second limitation: we don't use the DXY. We use the Federal Reserve's broad dollar index — freely available and complete, which is what it takes for anyone to recompute our result. It isn't the same thing. So we don't call it the same thing.

The four tests

One: is the sample large enough?

Ten trades. Our rule: below 30 trades, a backtest is an anecdote, not evidence. On very coarse timeframes we lower the bar, because fewer signals occur by construction — but we don't lower it to ten.

Ten trades across eight years is a travel diary.

The harder question sits underneath: can a two-week chart produce enough trades to establish anything at all? We're counting.

Two: does the denominator carry information?

Bitcoin divided by dollar sounds like macro analysis. But the dollar moves in percentages and Bitcoin moves in multiples. A divisor that barely budges, underneath a numerator that explodes, is cosmetic.

We run the same strategy on Bitcoin alone and on Bitcoin over dollar — and we compare the signal timestamps, not the returns. If more than 90 percent of signals are identical, the denominator is decoration and the macro framing is unearned.

That's our expectation. The threshold for "the denominator does contribute" sits below 75 percent overlap. If the test comes out the other way, we write that.

Three: where does the first candle start?

Two-week candles don't exist on any exchange. They have to be built from daily data — and that means deciding where the first one begins. There are two options.

Shift by one week and every candle in the chart changes. Different highs, different lows, different bands, different signals. With ten trades, one shifted starting point can flip the entire record.

This parameter appears in none of the charts you'll find online. It's a free value nobody declares.

We compute both. If the trade counts diverge by more than 20 percent, or the sign of the return flips, the result measures the choice of grid — not the market.

Four: can the signal see the future?

This is the decisive test, and it may explain the 100 percent entirely.

A two-week candle is unfinished for fourteen days. If the band calculation includes the live candle, the signal fires at that candle's close, and the trade executes at that same price — then the signal knows the high and the low of its own candle.

A signal like that cannot lose.

We're not claiming that's what happened. We're saying it's testable — and that a screenshot cannot answer it.

We compute both executions: once at the close of the same candle, once at the open of the next. If the win rate collapses by more than 15 percentage points, the first variant is contaminated. The 100 percent isn't a property of the strategy. It's a property of the bookkeeping.

If you're wondering how we check this mechanically: our test for look-ahead bias requires that a detector run over data up to day t produce the same signals as one run over the full history. If it doesn't, it's seeing the future.

The real test: 160 configurations

Four ATR lengths. Five multipliers. Two candle anchors. Two denominators. Two execution modes. That's 160 combinations.

We compute all of them. And we publish all of them.

Because this is what it comes down to: every screenshot online shows one cell. None of them show the grid.

If 160 configurations of the same system scatter across several orders of magnitude, then the one cell showing 34 million percent isn't a discovery. It's a draw.

That can't be argued away, and it requires no accusation. It only requires the spread.

This is the same principle behind how we evaluate strategies on the platform: not the best cell, but the distribution it came from.

What remains

If the system passes all four tests, we publish that. No hedging, no retroactive criticism, none of the tone people use when reporting results they didn't want.

We think it's unlikely. But that's an expectation, not a result — and expectations are exactly what pre-registration exists to constrain. So it's in writing, before we compute.

The result gets its own article. It will carry the numbers that are missing today — and the thresholds they have to be read against.

Frequently asked questions

What is overfitting in a backtest? Overfitting happens when parameters are tried until the result looks good. The winning combination then measures the search, not the market. You spot it when neighbouring configurations perform far worse.

What is look-ahead bias? Look-ahead bias occurs when a signal uses data that wasn't available at the moment of decision — for example the closing price of the candle it trades on. It produces win rates that are unreachable in practice.

How many trades does a meaningful backtest need? We set the threshold at 30 trades. Below that, a backtest is an anecdote. On coarse timeframes we lower the bar, because fewer signals occur by construction — but not without limit.

Why does a 100% win rate say so little? Because across ten trades it can arise from chance, overfitting or look-ahead bias. A win rate quoted without the trade count and the execution rules isn't a metric. It's a claim.

What does pre-registration mean? You state publicly, before computing, what counts as passing. That makes it impossible to fit the standard to the result afterwards.

Study the Past — Improve your Future. 🥋