Opening-range breakout — crypto

Measured in R: one R is whatever that day’s stop was risking. It has to be R and not points, because 10 points on Bitcoin and 10 points on SPY are wildly different bets. Above the dashed line is profit. Both lines take the same trades — they only differ on where profit is taken.

This is an invented account laid on top of real trades. The trades are real in the one way that counts — the entry, the stop and the target were written down before anyone knew how they ended. The dollars are invented. Nobody ever had $100 in this.

The rule. Risk 1% of the balance on every trade. A trade stopped out loses that 1%; a trade that makes twice what it risked makes 2%. That 1% is a number I chose, and it is the whole scale of the answer: at 0.5% the same trades move the balance half as far, at 2% twice as far. There is no slider, because on this little live data the relationship is exactly linear — a slider would let you make the number bigger without teaching you anything.

Fees are charged, and they decide the sign. Every other number on this site ignores costs, which is fine for a rule measured in R. A balance in dollars is a claim about an account, and accounts pay fees. The stop here is usually well under 1% of the price, so a 0.2% round trip eats roughly a quarter of an R every trade — against a gross edge nearer three hundredths of an R. Leave costs out and the number is not slightly rosy, it is the wrong sign.

Trades on the same day are added up, then applied once. Four coins broke out on the same morning. Compounding them one after another would pretend each later trade risked a balance that already held the earlier ones — across the full record that overstates the 2R version by 59%.

Vietnamese shorts are left out, because HOSE has no short selling and they were never placeable. Live calls only — the backfilled record is history the rule had already seen, and a dollar figure from that is money invented out of hindsight.

Win rate on its own is a trap. A fixed target that sits very close to the entry wins nearly every time and still loses money, because the stop is far away. Per trade is the number that decides it — and below about 0.15R it is indistinguishable from luck at this sample size.

These two rows are the same rule on the same four coins over the same days. The only thing that differs is what counts as the open. If the rule were picking up something real about how a market behaves after it opens, the choice of bell should not decide whether it makes money — and yet on the first 60 days backfilled, it decides the sign. One bell profits, the other loses, on identical prices. Read that as the size of the thing this rule is actually measuring: the hour I picked, more than the market. The US-open bell does catch a session about twice as wide as the UTC one, so 13:30 UTC is genuinely when crypto gets busy — but a rule that needs the right arbitrary constant to work has not been shown to work.

Two more things worth saying out loud. Crypto can be shorted, so unlike Vietnam every trade here is one you could place. And unlike the stock markets there are no holidays, no half days and no weekends off — which means roughly 40% more sessions per calendar month than the stock pages get.

The rule says “aim for 10 points”. Ten points is an ES number, so on everything else the target is the same size of move scaled to that instrument — about a quarter of its typical opening range.

Replayed over past data rather than watched live. Useful as a sanity check, worth less than the live number above.

The left half is not evidence. The right half is, because the search never saw it. Everything below is what happened when the first version of this rule was taken apart and rebuilt on data from the exchange itself.

What went wrong the first time. The earlier version was chosen out of 525 configurations and reported +0.048R per trade on the unsearched months. It was measured on a composite price feed — quotes stitched together from several exchanges. Every decision this rule makes turns on an exact touch: did the high reach the range edge, did the low reach the stop. A stitched feed's wicks reach further than any single venue's, so it reports levels being touched that were never touched anywhere an order could have rested. The same rule, same months, on Binance spot: +0.000R. On XRP the two feeds disagreed about the sign. The edge was the feed.

So the feed was replaced. Every crypto price here now comes from one exchange's order book. The backfill was rebuilt from scratch on it; the live record was not touched, because deleting live calls is the one thing this project will not do.

And the rule was rebuilt on top of it. Three changes, and only the last is a performance choice:

Ten coins, not four. Six of them were never in any search. This is the test that matters, and it is what exposed the old version: at a six-hour range the rule earns +0.127R on the four coins it was chosen from and −0.020R on the six it was not. It had been fitted to its own watchlist. At twelve hours the same split reads +0.082R and +0.060R — it survives coins it never saw.

A twelve-hour opening range, not six. This is arithmetic before it is a backtest. The fee is a fixed share of the price and the risk is the width of the range, so doubling the range halves what the fee costs in units of risk: 0.034R down to 0.025R.

Only when the opening range is wide — at least 1.2x the median of the previous twenty sessions. This was the only filter of the several tested that improved the typical configuration rather than the best one, on both halves of the data. It is worth being exact about why it works, because it is not what it looks like: breakouts from a NARROW range have a bigger gross move (+0.043R against +0.030R), but they pay 0.044R in fees to earn it while a wide one pays 0.017R. The filter is not picking better trades. It is declining to overpay for them.

What it is worth, and what it is not. Net of fees the filter turns −0.032R per trade into +0.035R. It survives every check that killed the last version: it holds on coins it was never chosen from, no single coin carries it (drop any one and it stays between +0.039R and +0.055R), and replaying it on one-minute bars gives the same answer as the hourly ones.

It is still not proven, and the reason is worth understanding. Counting every trade separately gives a t-statistic of +3.2, which looks decisive. But crypto moves as a bloc — on 42% of the days when more than one coin traded, every one of them went the same way. Ten coins breaking out together is closer to one observation than to ten. Clustered by day, which is what actually varies independently, the edge is +0.014R with a t of +0.60. That is indistinguishable from luck.

The honest summary: this is the best-evidenced version of the rule built so far, it is measured on the venue where the order would really rest, and it is not statistically significant. The live record below is what settles it, and nothing above can.

One question that is fully closed. On an hourly bar that reaches both edges you cannot tell whether you were filled and then stopped or stopped and then filled, and the rule refuses those. Seven months were replayed at one-minute resolution to check what that refusal was hiding: 0.6% of trades, worth six thousandths of an R. It was worth checking and it was not the problem.

Two new rules, and what was decided before the numbers came in. Both take the exact same trades as the 2R rule — same entry, same stop, same target. They differ only in which trades they will accept at all.

Cost-viable only refuses a trade when the round trip would cost more than 0.10R. That number means something precise: it is the amount fees inflate the hit rate you need, so 0.10 says “fees may make this rule at most 10% harder than it looks on paper”. It is a round convention, not a discovery — and it is not the best number on this data. The best is 0.145. Shipping the best-fitting constant is the exact mistake this site exists to warn about, so it is not what shipped.

With the trend refuses a trade taken against the average of the previous 20 session closes. Twenty sessions is the market’s conventional month, settled decades before this data existed. The trend for any session is built only from sessions that had already finished — the check that proves it is in orb.py --selftest, and it fails if the code is ever changed to let a day see its own close.

What the filters actually did. Cost-viability is not a filter outside the US — it is a shutdown: it keeps 212 of 214 US trades but only 3 of 411 in Vietnam and 2 of 240 on the UTC crypto bell, because those stops are never wide enough to carry their own fees. Vietnam’s is structural: the ±7% daily band caps the range, and 0.1% of the round trip is legislated tax.

Neither is better. On the backfilled record both come in behind plain 2R. No variant here will be called better or worse until an interval that accounts for how correlated these trades are excludes zero, and on this much data that will not happen. “Not enough evidence, and here is how many trades it would take” is the claim.

This is why the rule loses, and it is not the exit target. The gross edge is real but tiny. The fee to get in and out is several times larger than it, because the stop sits well under two percent of the price while a round trip costs a fifth to half a percent. You are paying a quarter of your risk for the privilege of taking a bet worth a few hundredths of it.

Win rate needed is what this payoff must achieve just to break even, before any costs — average loss divided by the sum of the average win and the average loss. Win rate got is what actually happened. The gap between them is the entire edge, and the fee column swallows it whole in every market.

The fee model is deliberately optimistic: 0.02% round trip on US instruments, 0.2% on crypto, 0.5% in Vietnam where brokerage on both sides is joined by a 0.1% tax on every sale. Real fills would be worse. No amount of moving the profit target fixes a gap this shape — it would take a wider stop, a cheaper market, or an edge several times bigger.

Written by the agent as it ran, not composed for this page afterwards. Fetches, ranges, breakouts and the runs that found nothing. The newest scheduled run is at the bottom.

Showing the latest 60 of 8396 — the full history lives in journal.json.