H3 — Flow-Confirmed Breakout Continuation: Findings & Recommendation

Status: complete — negative result.

Design: reports/cvd_breakout_flow_study.md (frozen before results).

Machine outputs: experiments/reports/breakouts/ and breakouts_all/ (summary.json, by_symbol.json, leave_one_out.json, quintiles.json, robustness.json, events.parquet).

Reproduce:

python scripts/breakout_study.py                    # 62 crypto perps (primary)
python scripts/breakout_study.py --include-noncrypto # 100 symbols incl. TradFi

1. Verdict

NO EDGE (H0) in both universes. All preregistered criteria R1–R7 fail. Persistent same-direction aggressive flow does not improve breakout continuation after accounting for the breakout itself.

2. What was tested

Close-beyond-prior-range breakouts (trailing L=48 extreme, excludes current bar), 1h bars, aligned activity-normalised flow over w=4 bars (flow_align_z), 24-bar cooldown, 4h primary horizon; outcomes = oriented forward return, continuation-failure rate, MFE/MAE; baselines = price-only breakout, momentum-only, and the prior study's cvd_20>0 filter.

3. Results

• Horizons (crypto-62 test IC): 1h −0.0179 · 4h −0.0146 · 24h −0.0115 · 3d −0.0086 — negative everywhere.
• Matched controls and the incremental regression both include zero; the regression point estimate is negative in both universes.
• Robustness (27 cells): at 15m/1h design and test IC are consistently negative; at 4h the design IC is strongly negative (−0.15…−0.21) while test flips small-positive — instability, not a stable region.
• Failure rate shows a mild decline with flow alignment (0.588 → 0.441 across quintiles) but forward returns are non-monotone and inverted — no tradable structure.

4. Relationship to the prior breakout study

reports/breakout_regime_study.md found the cvd_20>0 breakout filter insignificant (validate lift 1.07). H3's result is directionally consistent: flow confirmation does not rescue breakouts, and where anything is detected it is slightly negative intraday. Four flow-conditioning ideas now fail on the same 2026 sample (divergence, H1 absorption, H2 sweeps, H3 breakouts).

5. Limitations

• Single regime, in-sample 2026 — no cross-regime out-of-sample.
• 1h resolution, taker-flow proxy, event-level samples of a few thousand to ~7k.
• Overlapping events and clusters handled by cooldown + timestamp-block bootstrap, but serial dependence remains a caveat.

6. Recommendation

1. Do not re-tune H3. No stable cell; both universes, all horizons, matched controls and the incremental regression agree on "no edge".
2. The recurring blocker is out-of-sample history. Four hypotheses have now been judged on a single 2026 regime. Building 2024–2025 perp 1m data (~2 h, ~3 GB) is the single highest-value action before any further flow hypothesis; without it, negative results are as unconfirmable as positives.
3. Consider stepping back from bar-level taker-flow CVD. Four independent flow features (spot-vs-perp divergence, impact residual, sweep-aligned flow, breakout flow persistence) show no incremental value at 1m→1h in this sample. Before more variants, either (a) raise the information content (tick/aggTrades for true sub-bar impact — but this is costly and lowers breadth), or (b) move to a structurally different information source (funding/basis carry, cross-sectional relative strength, liquidation-adjacent proxies where data exists).
4. If continuing the H-series: H6 (spot lead–lag) is the last genuinely distinct flow test but risks re-treading the failed divergence study; H5 (impact asymmetry) is better kept as an execution/cost model than a signal.
5. Do not buy tick data yet; the binding constraints are regime breadth and effect existence, not resolution.

Bottom line: abandon H3; prioritise out-of-sample history, then question whether 1m-bar taker-flow CVD carries incremental information at all.

← All reports · source (.md)