Pattern Reliability During High Volatility

Methodology

Every element below, including the regime cut-offs, is registered before outcomes are computed, so the definition of high volatility cannot be tuned to the results.

1. Regimes from the past only. At each signal bar, volatility is ATR(14) as a percentage of price, the measure PatternX AI's volatility-regime factor reads. The regime is its percentile rank among the same series' previous 500 completed bars: high above the 80th percentile, low below the 20th, normal in between. A later crash can never reclassify an earlier detection. A market-wide variant ranks BTC's ATR% on the same timeframe the same way, separating crypto-wide stress from coin-specific moves.
2. Absolute bands. The PatternX Score gives its volatility factor most credit for ATR% between 0.2% and 2%, so outcomes are also reported below, inside and above that band.
3. Replay. All 20 detectors run on completed bars only, with the 200-bar live window; candle-based filters apply as in production and live-only filters are neutral. Entry is at the next bar's open, levels are frozen at detection, a bar touching both levels counts as a stop, and gap-throughs fill at the open.
4. Regime-matched controls. Ten random entries per detection on the same series, in the same regime, month and direction, with the same level distances. Volatility changes how quickly any bracket resolves, so regimes are compared on excess R, not raw hit rates.
5. Pattern mix. Most detector thresholds are fixed percentages, so volatility changes which formations qualify at all. Pooled results are standardised to the normal regime's pattern mix, and every pattern is also reported on its own.
6. Costs. Base case: 0.03% slippage each way, 0.05% fees per side, 0.01% funding per 8 hours on perpetuals. Stress case: five times the base slippage in the high regime, because fills in fast markets are the weakest part of any simulation built on OHLC bars.
7. Statistics. The confirmatory contrast, run once on the out-of-sample last 30% of the period, is mean excess R in the high regime minus the normal regime; low versus normal is secondary, and both are tested two-sided at the 5% level after Holm adjustment. High-volatility bars bunch into a few episodes, so intervals come from a bootstrap that resamples calendar weeks. Hit rates carry 95% Wilson intervals, and each regime's full R distribution is published, including its worst 1% of outcomes. Per-pattern, per-market and per-year results are exploratory.
8. Reproducibility. The detector commit and a hash covering the 500-bar lookback, percentile cut-offs, filters and costs are published with the data checksums and seeds.

How success is defined

Primary outcome. Whether each detection reaches its target before its stop within H bars of entry, and its net R-multiple. H equals the detector's trailing window: 40 bars for flags and pennants, 45 for rounding patterns, and 70 to 110 for the rest. Target and stop are the detector's own levels, with production's 0.2% stop buffer, frozen at the signal. A bar spanning both counts as a stop; reaching neither within H bars is a time exit and a non-success.

Compared quantity. Excess R, a detection's net R minus the mean net R of its regime-matched random entries. The confirmatory contrast is mean excess R in the high regime minus that in the normal regime, with a 95% week-block bootstrap interval.

Tails, not just averages. Fast markets can gap through stops, so each regime reports the share of trades losing more than 1.5R, its worst 1% of results and its share of time exits.

Secondary outcomes.

• Production defaults by regime: the share of detections whose stop would be tightened to the 5% default, whose target would be capped at 10%, and that would be dropped for reward:risk below 1.
• Market-wide versus coin-specific stress: the same contrast using the BTC-based regime.
• Excursions and speed: maximum favourable and adverse excursion in R, and bars to resolution, per regime.

Sensitivity. Percentile cut-offs at 10/90 and 25/75; a 250-bar lookback; the slippage stress case.

Pattern and strategy criteria

All 20 production detectors run unchanged; each pattern's complete rules are in the pattern library. Because most thresholds are fixed percentages of price, volatility changes which formations qualify. The thresholds most exposed to that effect:

• Universal filters: the breakout candle's close in the trade direction; breakout volume ≥ 1.2× the 20-bar average; ATR% ≥ 0.05, a floor only quiet bars can fail; liquidity score ≥ 30; spread ≤ 0.25%; order-book imbalance no worse than −0.5; crypto 24-hour volume ≥ $3M. In replay the first three apply and the rest are neutral.
• Levels: pattern stops and measured-move targets as each detector returns them; reward:risk below 1 dropped.
• Volatility after detection: apart from the ATR% floor, no detection rule measures volatility directly. Volatility enters the PatternX Score, and the optional AI context layer, only after a setup has been found; neither is used to filter detections in this study.

Exclusions

• Too little history to rank: detections without 500 completed bars before the signal, in addition to the 200-bar detector window.
• Isolated bad prints: bars whose high or low lies more than 25% beyond the closing prices of both neighbouring bars, a threshold borrowed from the position manager's bad-tick rule. Each such exclusion is counted by regime, because genuine extreme moves must stay in the sample.
• Venue outages and trading suspensions: detections whose outcome window contains one, counted by regime.
• Thin crypto markets: trailing 24-hour candle volume, valued at closing prices, below $3M; stablecoins and pegged assets; a contract's first 7 days after listing.
• Stocks: outcome windows containing a trading halt; corporate actions the data does not adjust for; bars outside the regular session.
• Duplicates and overlap: same-direction detections on one series and bar are merged; one open trade per series and pattern.
• Untradeable or censored: detections whose next bar opens beyond a level, and those with fewer than H bars of data left. Both are counted.

Limitations

• Exclusions cluster in stress. Outages, halts and the bad-print rule are most likely to remove detections from the high regime, which can bias it; the counts are published so readers can judge the effect.
• OHLC bars hide fill quality. During liquidation cascades a stop can fill far beyond its level. The five-times slippage case is an assumption, not a measurement.
• Few independent episodes. High-volatility periods come in a small number of clusters, so the effective sample is closer to the number of episodes than to the number of bars.
• Relative regimes. A percentile regime is relative to each series' own past: high for BTC and high for a small coin are different absolute conditions. The absolute ATR% bands address this only partly.
• Standardisation is partial. Re-weighting to the normal regime's pattern mix removes differences in mix between patterns, not differences in geometry within a pattern.
• Live risk controls are absent. In live use, the daily loss limit, the 60-minute cooldown after three losses in a row and the per-symbol bench after five losing trades in a day can stop trading during a volatile stretch. The replay keeps every detection, so it overstates how many high-volatility trades a live account would take.
• Neutral context and timing. Spread, order-book and funding data are absent; forming bars, scan intervals and approval delays are not modelled. Engine-wide caveats are collected on the limitations page.
• Regime dependence by construction. Future volatility episodes may differ in cause and shape from those in the sample period.