Breakout, Confluence & Regime

§ 1Research question and thesis

The lab’s reference strategy is a volatility/band breakout: price crosses above a volatility band on expanding volume. The working hypothesis was:

H1. Raw breakouts are noise, but a confluence of conditions — volume, order-flow, trend, volatility regime, derivatives — selects the subset that trends. Working hypothesis

A second hypothesis followed from the data:

H2. The edge is regime-dependent — breakouts work in trending/expanding regimes and fail in range/contraction regimes. A regime classifier should therefore separate signal from noise. Follow-on hypothesis

Both were tested on a design period and a separate validation period, with a pre-registered outcome definition, and H2 was additionally stress-tested by replicating 2026 on 2025.

§ 2Data

2.1 Sources and coverage

All data from Binance Public Data bulk archives (data.binance.vision), via a resumable, checksum-verified downloader (quant_lab.data.sources.binance).

Ranges: 2025-01-01 → 2025-12-31 and 2026-01-01 → 2026-08-31. Also present: BTCUSDT/ETHUSDT futures 2024-01..03, and 2024-01/02 funding plus five days of OI.

2.2 Storage and validation

Canonical schemas (quant_lab.data.schemas), Parquet partitioned by exchange/market_type/dataset/symbol/timeframe/year/month.parquet, queried via DuckDB. Validation (quant_lab.data.validate) checks missing, duplicate and non-monotonic timestamps, invalid prices and quantities, gaps, outliers and timezone consistency.

Every dataset used here passed validation cleanly. No data was fabricated or silently repaired.

2.3 Known data caveats

• Monthly archives cover whole calendar months; query-time filtering bounds use.
• Funding timestamps carry ~1–15 ms jitter; truncated to the minute for joins.²
• open_interest is derived from the daily metrics archive (5-minute grid).
• Liquidations are unavailable and were not used.
• Trade-level data was not downloaded; 1m bars suffice for the flow features used.

§ 3Methodology

3.1 Triggers tested

All strategies are long-only by default, use trailing (causal) windows only, and are validated for causality by tests that assert signal and feature values do not change when future bars are appended.

3.2 Execution assumptions

Recorded with every run; defaults are non-zero.

3.3 Outcome measures

For every signal, the forward path is measured from the close of the signal bar over a 7-day horizon:

time to reversion

first bar whose close returns to the signal price (a stricter “touch” variant using the bar low was also computed).

excursion to reversion

the maximum favourable price move (highest high for longs) between the signal and the reversion bar.

good outcome (label)

excursion ≥ X && time_to_reversion > 1h, at X = 0.25% (1m study) and X = 0.5% (main study). A “never reverted” signal counts as slow.

Reversion and excursion are also reported as a cumulative reversion curve and a bucketed excursion table (disjoint reversion-speed buckets).

3.4 Confluence filters

Each filter is a single explicit boolean condition evaluated at the signal bar (no model): volume ≥ 1.5×mean, taker_buy/volume ≥ 0.55, 20-bar CVD > 0, close > EMA200, strong close (close in top 30% of bar range), bandwidth z ≤ −0.5, spot–perp basis > 0, funding ≤ 0, OI 1h change > 0. Combinations are greedy ANDs of the best standalone filters on the design period.

3.5 Regime classifier

A causal classifier (quant_lab.analytics.regimes), computed on a 1h regime timeframe from trailing statistics only:

• trend — up/down/range from a volatility-normalised trailing move: the 24h log return divided by realized_vol × √24, threshold ±1.0.
• volatility — highvol/lowvol from realized vol vs its trailing median.
• combined label {trend}_{vol} (e.g. up_highvol), shifted one regime bar so a signal only ever sees regimes that had already closed.

3.6 Evaluation protocol

• Design / validation split. Filters and the regime hypothesis were chosen on a design window and reported on a disjoint validation window. The final test used design = 2025 and validation = 2026 Jan–Aug.
• Baseline and lift. For each filter, lift = P(good | filter) / P(good | baseline).
• Significance. Two-proportion z-test of the filtered rate against the baseline rate (2-sided p-value) — essential because the base rate is low and samples are small.
• Multiple testing. Many filters × periods were tested; interpreted with that in mind. No correction is claimed, so only large, replicated effects would be credible.

§ 4Results

4.1 Phase 1 — Reversion and excursion at 1m

2,492 Bollinger breakout signals, 1m, close-basis reversion, 7-day horizon. Median time to reversion was 1 bar (1 min), mean 51.2 min, p75 4 min; 21 signals (0.8%) never reverted within 7 days.

4.2 Phase 2 — Confluence at 1m

Design = July, validate = August, 0.25% label. Baseline P(good): 6.28% design, 6.16% validate.

4.3 Phase 3 — Higher timeframes

Jul/Aug split, 0.5% label. The base rate rises roughly 4× (15m) to 6× (1h) versus 1m — but signal counts collapse and range_breakout_hold is too restrictive to evaluate. No filter was significant at any timeframe.

4.4 Phase 4 — Full 2026, design Jan–Jul / validate Aug

At 15m, two filters were positive in both windows: volume ≥ 1.5× (1.13/1.14) and close > EMA200 (1.11/1.15); their combination gave 1.19/1.23 — the first directionally consistent signal, though not significant. Adding oi_change_1h produced a large validation lift (2.05, n=20, z=1.94) but its design lift was only ~1.06 — treated as regime-specific, not real. At 1h no filter helped; breakouts there are already good about a third of the time.

4.5 Phase 5 — Regime classifier, 2026

Baseline: 17.7% design (Jan–Jul), 19.5% validate (Aug). up_highvol was the best regime in both windows, and down-regimes were worst. At this point H2 looked supported — pending the 2025 replication.

4.6 Phase 6 — 2025 replication

Baseline: 18.8% (2025), 17.8% (2026 Jan–Aug).

Regime-gated backtest

Gate = up_highvol + up_lowvol.

§ 5Interpretation — why it fails

1. The trigger has no edge. A band/volatility cross fires on a large fraction of bars and, at 1m, is dominated by microstructure; at 15m/1h it is better but still a coin-flip biased toward fast reversion.
2. The “good” outcome and PnL diverge. “Excursion ≥ 0.5% before reversion” is not a profit measure: the strategy exits at the mid-band, pays fees on both sides, and gives back the excursion. A signal can be labelled good and still lose money.
3. Regime effects are not stable. The apparently strong 2026 up_highvol result is a single-sample artifact; 2025 contradicts it. With ~60–80 signals per regime-year, confidence intervals are wide enough to manufacture such effects by chance.
4. Low base rates destroy statistical power. At 1m the good rate is ~5% with thousands of signals; at 1h it is ~⅓ with only ~25/month. There is no window where both the effect size and the sample are large enough to prove a modest (1.2–1.3×) lift.
5. Gating’s apparent benefit is a turnover effect. Fewer trades → fewer fees → smaller losses. That is risk reduction, not alpha.

§ 6Limitations

• Single instrument (BTCUSDT). No cross-sectional validation.
• Two calendar blocks (2025, 2026 Jan–Aug) with different regime mixes; neither is a large sample at higher timeframes.
• Multiple comparisons. Many filters × timeframes × periods were tested; no formal correction applied. Any single small lift should be discounted.
• Outcome definition is a proxy. The good label is path-based, not PnL-based; it ignores the exit rule, funding and slippage on exits.
• Backtest simplifications. Bar-based fills, fixed per-side slippage, intrabar stop/target ambiguity handled by an explicit policy, no order-book or latency modelling, funding applied as a periodic drag.
• Regime thresholds are hand-set (24h trend z, ±1.0; vol vs trailing median) and were not themselves optimised or cross-validated.
• Derivatives history is short relative to bars; OI/funding filters were regime-dependent and sometimes empty (e.g. no non-positive funding in Aug 2026; no positive basis in Jul 2026).

§ 7Conclusion

The hypothesis that a breakout trigger plus confluence or a regime gate yields a durable BTCUSDT edge is not supported by 2025–2026 data.

• Timeframe is the only robust lever: it raises the “good” base rate from ~5% (1m) to ~18–20% (15m) to ~32–36% (1h), consistently.
• No confluence filter reached significance at any timeframe; the best candidate (volume + EMA200 at 15m) was directionally consistent but weak (1.23× out-of-sample, z≈0.6).
• The regime hypothesis failed out-of-sample: up_highvol was special only in 2026; in 2025 a down regime was best.
• Regime gating reduced losses mainly by cutting turnover; it did not make the strategy profitable.

§ 8Reproducibility

Data

quantlab data build binance --market futures --symbol BTCUSDT --interval 1m \
    --start 2025-01-01 --end 2026-01-01
quantlab data build binance --market futures --symbol BTCUSDT --dataset funding \
    --start 2025-01-01 --end 2026-01-01
quantlab data build binance --market futures --symbol BTCUSDT --dataset open_interest \
    --start 2025-01-01 --end 2026-01-01
quantlab data build binance --market spot --symbol BTCUSDT --interval 1m \
    --start 2025-01-01 --end 2026-01-01
# (repeat for 2026-01-01 → 2026-09-01)
quantlab data validate --symbol BTCUSDT --dataset bars --interval 1m

Reversion / excursion study

quantlab analyze excursions --strategy bollinger_breakout --symbol BTCUSDT \
    --timeframe 1m --start 2026-07-01 --end 2026-09-01 \
    --reversion-only --horizon-days 7 --reversion-basis close \
    --out experiments/reports/bb_reversion_only

Confluence and regime evaluation

quantlab analyze confluence --strategy bollinger_breakout --symbol BTCUSDT \
    --timeframe 15m --source-timeframe 1m \
    --start 2025-01-01 --end 2026-01-01 \
    --validate-start 2026-01-01 --validate-end 2026-09-01 \
    --excursion-pct 0.005 --min-reversion-hours 1 --horizon-days 7 \
    --by-regime --regime-timeframe 1h \
    --out experiments/reports/regime_2025_2026

Regime-gated backtest

quantlab analyze regime-backtest --strategy bollinger_breakout --symbol BTCUSDT \
    --timeframe 15m --start 2025-01-01 --end 2026-01-01 \
    --allow-regimes up_highvol,up_lowvol \
    --out experiments/reports/regime_bt_2025

Key artifacts

Code map

Tests

uv run pytest — 97 tests, including deterministic end-to-end, lookahead/causality, resampling, regime and gating tests.