ATR Volatility Breakout on EUR/USD 4H: 27 Configurations, Zero Survivors

1. Hypothesis & Parameter Grid

The hypothesis: when price closes decisively beyond its recent range — scaled by current volatility so that "decisively" means the same thing in quiet and turbulent markets — the move tends to continue for at least a few bars. ATR (Average True Range) provides the volatility scaling.

Signal type: Volatility breakout, long and short — the first bidirectional study in this series. Entry long: close exceeds the prior N-bar high by more than mult × ATR(14). Entry short: close falls below the prior N-bar low by more than mult × ATR(14). Exit: fixed number of bars after entry. One position at a time; signals during an open position are ignored.

Grid structure: full factorial, 3 × 3 × 3 = 27 configurations (hypotheses HYP_0157 through HYP_0183). Transaction cost: 2.0 pips round-trip.

Note what this definition demands: not merely a new high, but a close a full 1.5 to 2.5 ATR beyond the range extreme. On EUR/USD 4H that is a 25 to 60 pip margin past a multi-day high or low. These are genuinely rare events — between 0.1% and 0.9% of bars qualify, depending on settings — and that rarity turns out to matter as much as anything else in this study.

2. Methodology

Identical methodology to Part 4 (Bollinger Bands) and every prior study: 13.3 years of EUR/USD 4H data (22,193 bars, January 2013 through April 2026), six non-overlapping walk-forward windows (2 years in-sample, 1 year out-of-sample), 2.0 pip round-trip costs, and the same staged kill filter — net expectancy, effective sample size, per-window IS/OOS win rate drift, binomial test — applied in order.

One methodological addition: because breakout logic is inherently two-sided, this is the first study in the series where the simulation opens genuine short positions. Long entries buy upward breakouts; short entries sell downward ones. Both directions face the same costs and the same filter.

All figures below are extracted programmatically from the stored sweep results. Every configuration produced a verified, distinct result set, with one documented exception covered in Section 7.

3. Results — All 27 Configurations

Kill funnel: 23 of 27 (85.2%) killed by negative net expectancy. 4 of 27 (14.8%) killed by insufficient effective sample size. Zero killed by IS/OOS drift. Zero survivors. Trade counts range from 95 (the most permissive settings) down to 12 — roughly 16 trades per year at the high end, 2 per year at the low end.

4. A Different Failure: Zero Drift Kills

Every previous family in this series was dominated by the same executioner. SMA, MACD, RSI, and Bollinger Bands all produced configurations that looked plausible in-sample and diverged out-of-sample — the signature of strategies fitting themselves to regimes that then ended. The per-window drift check was the busiest stage of the kill filter for four consecutive studies.

The ATR breakout is the first family where that stage never fired. Not one of the 27 configurations was killed by drift. There was nothing for drift to catch, because the strategies never looked good enough in-sample to have anything to lose out-of-sample. They lost money in-sample, they lost money out-of-sample, they lost money in trending years and ranging years, long and short, at every multiplier.

This is worth sitting with, because it is a fundamentally different — and in a sense more useful — verdict than the previous four. Regime overfitting says: your backtest lied to you. Negative expectancy everywhere says: the market itself charges for this behavior. A 25-to-60-pip move beyond a multi-day extreme on EUR/USD 4H is not the beginning of something. On average, it is the end of something — and the trader who buys it is providing exit liquidity to whoever rode the move there.

The per-trade losses are not subtle. The most active configuration (lookback 10, multiplier 1.5, 8-bar exit) averaged -18.71 pips per trade across 93 trades. Costs are 2 pips of that. The other 16.7 pips are the market taking back the breakout.

5. The Positive Cells Are Statistical Ghosts

Four configurations show positive net expectancy, and it would be easy to squint at the table and see a tradable island: lookback 20, multiplier 2.0-2.5. The kill filter's sample-size stage exists precisely for this squint.

The best of them, HYP_0169 (lookback 20, multiplier 2.0, 3-bar exit): +0.45 aggregate Sharpe, +2.53 pips per trade — on 28 trades in six out-of-sample years. Its per-window record shows what an average over tiny samples conceals:

One window contains two trades. Another contains one. The best window (+25.95 pips average) is an average of exactly two events. A single 2015 window with nine trades erased a quarter of the six-year total by itself. There is no statistics to be done here — only anecdotes with decimal points.

The pattern repeats at HYP_0174 and HYP_0183 (+6.07 pips, 12 trades each, roughly two per year). The grid's only positive cells are exactly its emptiest ones. That is what selection over noise looks like: make events rare enough, and some corner of the grid will always be lucky.

6. Holding Longer Loses More

If breakouts began sustained moves, longer holding periods should capture more of them. The data says the opposite:

Whatever follow-through exists at the moment of an extreme close is spent within the first few bars. Give the position more time and you are not giving the move room to continue — you are giving it room to come back. The fastest exit tested was the least bad at every lookback and multiplier. For a momentum event, that is damning: momentum that decays inside 12 hours is not momentum a 4H swing trader can use.

7. A Resolution Limit of Rare-Event Backtests

One honest wrinkle from the audit, published here because anyone backtesting rare events will hit it eventually. At the strictest multiplier (2.5), the lookback-20 and lookback-30 configurations produce identical out-of-sample results for every exit setting — three pairs of duplicate outcomes in the table above.

The parameters are not broken, and the configurations are not the same strategy. A close that clears the 30-bar extreme by 2.5 ATR nearly always clears the 20-bar extreme too, so the stricter signal is a subset of the looser one. Over 13.3 years the two differ by exactly four entries — and every one of those four happens to fall outside the six out-of-sample evaluation windows. Within the windows the evidence can see, the two strategies did precisely the same things.

The lesson generalizes: when a strategy trades a dozen times in six years, a walk-forward backtest cannot distinguish it from its near neighbors, cannot estimate its edge, and cannot be refuted by its failures or vindicated by its successes. The sample-size stage of the kill filter is not bureaucracy. It is the boundary between measurement and folklore.

8. Five Families Compared

Running total: 147 configurations tested, zero survivors. But the failures are not interchangeable, and the differences are the real product of this series. Pattern indicators fail by promising an edge in-sample and withdrawing it out-of-sample. The breakout fails by never promising anything at all — and in doing so, it reveals something about the market the others only hinted at.

9. What This Market Is Telling Us

Put Parts 4 and 5 side by side and a consistent picture of EUR/USD 4H emerges. The mean-reversion family (Bollinger Bands) produced the best raw per-trade economics in the entire series — buying statistical extremes and betting on a snap-back. The breakout family, which bets the exact opposite way at similar extremes, loses money almost everywhere, in both directions.

Same market. Same extremes. Opposite bets. One side of the trade consistently holds the better cards — and it is the fading side, not the chasing side. EUR/USD at the 4-hour scale behaves like a market that punishes momentum at its extremes: moves that have already traveled 1.5 to 2.5 ATR beyond their recent range are, on average, closer to their end than their beginning.

None of this rose to a deployable edge — the fading side's advantage was regime-dependent, and the chasing side's disadvantage is not directly tradable after costs. But as a structural fact about the instrument, established from two independent directions with 36 configurations, it is exactly the kind of raw material the combination phase of this research program needs: not "which indicator wins," but which way this market leans, and when.

Five families in, the series' central claim stands unbroken and better-evidenced than ever: on EUR/USD 4H, under honest walk-forward validation with real costs, no single indicator with fixed parameters survives. If someone sells you one that does, ask to see their per-window table — and their trade counts.

10. Frequently Asked Questions

11. Next in the Series

This is Part 5 of the MT5 Indicator Backtest Series. Every study uses the same methodology, the same dataset, and the same staged kill filter, making results directly comparable across indicator families.