Opening Range Breakout Failure Rate: What 16 Years of ES Data Show
A breakout can look clean and still close back through its level a few bars later. Across roughly 3,900 qualifying ES events from 2010–2026, that happened 52.1% of the time within 25 minutes and 70.6% within 100 minutes. The definition matters: this is a rolling 10-bar range, not the usual fixed 5-, 15- or 30-minute ORB.
Quick answer
Among qualifying prior-10-bar ES breakouts, a later close crossed back through the breakout level in 52.1% of events within five bars, 62.6% within ten and 70.6% within twenty. The horizons are cumulative: an event that fails in bar four also counts as failed by bars ten and twenty.
These percentages are historical event outcomes, not a trading system’s loss rate, a fade strategy’s win rate or proof that every opening-range breakout fails.
The exact breakout and failure definition
- Before each candidate five-minute bar, take the high and low of the immediately preceding 10 bars.
- Keep the range only when its width is no greater than 0.5 times ADR5. ADR5 is the mean high–low range of up to five prior complete RTH sessions; fewer than three prior sessions means no read.
- A breakout requires the current bar to close strictly above the range high or strictly below the range low. A wick through the boundary is not enough.
- Keep only the first qualifying event in each window and day.
- A failure occurs when a later bar closes strictly back through the original breakout boundary. Equality is not a failure, and price does not need to cross the far side of the range.
That last distinction prevents a common misquote. This study asks whether the close returns through one breakout level. It does not ask whether price traverses the entire range.
Measured ES rolling-range failure rates
| Follow-up horizon | Elapsed time | Failed by horizon | Eligible events |
|---|---|---|---|
| 5 bars | 25 minutes | 52.1% | 3,925 |
| 10 bars | 50 minutes | 62.6% | 3,922 |
| 20 bars | 100 minutes | 70.6% | 3,915 |
Sample counts fall slightly at longer horizons because an event needs enough later bars inside the same session to be scored.
Why this is not the fixed opening-range result
MADDOG’s broader opening range breakout strategy study also reports an archived 83.5% result. That experiment forms one fixed range from the first 18 bars and asks whether the opposite session extreme survives after the first later closing break.
The study on this page continuously rolls a 10-bar range forward until the first qualifying event, then looks for a close back through the breakout boundary. Different range, event, outcome and sample: 83.5% is neither the success rate nor the complement of 70.6%.
What the rates do — and do not — imply
The useful finding is about path, not profit. Re-entry through the breakout level is common and keeps accumulating with time. A trader testing continuation should therefore state how much follow-through is required and when the setup expires; a trader testing a fade should still specify entry, stop and exit instead of treating 70.6% as an automatic win probability.
No fees, slippage, position sizing or P&L rule appears in this measurement. A close back through a level can happen after a large favorable excursion, and an event that never closes back can still be untradeable under a particular stop. The table cannot settle either strategy without those missing rules.
Methodology and evidence limits
The definition was recovered after the original aggregate rates were already known, then frozen before cross-window use. These tuples are stable production references, not a first-look, pre-registered replication. The source bars and a one-command generator for the exact aggregate table are not distributed on this public site.
The historical research grid contains 81 bars through 16:15 ET; the current live RTH endpoint stops at 16:00 ET. Late events without enough follow-up bars are excluded, but the windows are still not literally identical. The published aggregate is an ES-specific statement. On NQ, individual stocks or crypto it is context, not a measured claim about that instrument.
A separate conditional model asks whether one event’s own context changes its ten-bar failure probability. Its original ADR20 model and NQ test protocol were specified before the NQ result; the production ADR5 port was then checked on NQ 2023+ at AUC 0.646, calibration error 5.1%, n = 2,637. That is model validation, not a validation of the aggregate table as a trading rule.
See the full methodology and provenance record and the machine-readable appendix.