CVD & Order-Flow Research Report

Headline

NO EDGE (H0)  All four order-flow hypotheses tested today were rejected. Every preregistered pass/fail criterion failed in every study, and no parameter region in any robustness sweep was stable.

Data covers 2026-01-01 → 2026-09-30 only for the tested universe (BTCUSDT is the sole symbol with multi-year history). Every verdict is therefore exploratory: a negative is as unconfirmable out-of-sample as a positive would have been.

Data & features

• Source: Binance public bulk archives; 1-minute OHLCV klines with taker-buy volume and taker-buy quote volume.
• CVD / flow: delta = 2·taker_buy_volume − volume, quote_delta = 2·taker_buy_quote_volume − quote_volume. Working in quote (USDT) units for cross-asset comparability. CVD is taker flow only — it cannot see maker intent, hidden liquidity, or cancellations.
• Resampling: canonical 1m→1h aggregation (open=first, high/low=extreme, close=last, volumes & taker-buy summed).
• Regimes: causal trend×vol classifier (up/down/range × high/lowvol), shifted one bar.

Causality & timing contract

• Every feature uses trailing windows only; prior-range levels exclude the current bar.
• Signals evaluated at the close of bar t; all outcomes begin at t+1 (one-bar execution delay).
• Explicit unit tests assert look-ahead freedom (appending future bars must not change past feature values) and forward-return alignment (no same-bar leakage).

Evaluation

• Pooled rank IC of the signal vs the oriented forward return, with block-bootstrap CIs (blocks ≈ horizon, to respect overlapping windows).
• Matched controls (high vs low signal, matched on volatility, trend, activity/liquidity and regime) and a control regression clustered by symbol×day.
• Chronological design (Jan–Apr) / test (May–Sep) split; full parameter sweeps reported; multiple-testing acknowledged (Benjamini–Hochberg in principle).
• Costs: ~12 bps round-trip (4 bps fee + 2 bps slippage per side).

Universal caveat: single regime (2026), in-sample. No cross-regime out-of-sample exists for this universe.

Question. Does divergence between spot taker flow and perp taker flow predict forward returns? A move led by spot may be “real” demand; a move led by perps may be crowded leverage prone to reversal.

Signal. D = z(spot rolling flow) − z(perp rolling flow), each flow a 60-bar rolling sum z-scored over a 720-bar window. Events = |D| ≥ trailing q95. Predicted: sign(D) → forward return (spot-led ⇒ continuation).

Data / spec. Top-10 crypto cross-market (present in both spot and perp top-100); 1h bars (from 1m); primary horizon 4h; market-neutral (cross-sectionally demeaned) forward returns; 12 bps.

Primary result — pooled rank IC(D, forward return)

Test 95% block-bootstrap CI: [−0.0239, +0.0170] (spans zero). Design IC crossed the 0.02 bar and then flipped sign out-of-sample.

Disconfirming evidence

• Deciles non-monotone (spearman −0.067); top−bottom decile spread −11.84 bps.
• Leave-one-out: dropping any single symbol leaves IC negative (−0.0011 … −0.0070).
• Ex-BTC/ETH IC: −0.0068.
• No criterion (C1–C7) passed; controls (C4) were moot.

Artifacts: reports/spot_cvd_divergence_study.md, reports/spot_cvd_divergence_results.md.

Question. When aggressive selling produces unusually little downward price movement (absorption), does price subsequently outperform comparable strong-selling events? Symmetric for buying.

Signal. Rolling OLS r ~ α + β·f over a trailing 720-bar window (f = quote flow / trailing activity). Residual s = r − r̂ (impact surprise); absorption score A = −sign(f)·s, z-scored. Strong-flow events = |f| ≥ trailing q90. Outcome = oriented forward return y = −sign(f)·fwd_ret. A regression-based impact model is used deliberately instead of a naive return/flow ratio (unstable as flow → 0).

Data / spec. Top-10 crypto perps; 1h; primary horizon 4h; 12 bps.

Primary result

Test CI [−0.0379, +0.0007]; strong-flow-only test IC −0.0047. Horizons: 1h −0.0171 · 4h −0.0187 · 24h +0.0014 · 3d −0.0035.

Artifacts: reports/cvd_absorption_study.md, reports/cvd_absorption_results.md, reports/cvd_absorption_findings.md.

Question. When price sweeps a prior swing level and closes back inside (a failed sweep / stop-run), does the magnitude and direction of aggressive flow during the sweep predict reversal strength?

Signal. Level = trailing prior-range extreme over 48 bars (excludes the current bar). Failed sweep = wick beyond the level, close back inside (24-bar cooldown). Aligned flow af = −d·F / activity (spends in the sweep direction), z-scored; outcome = oriented reversal return y = d·fwd_ret. Hypothesis: large aligned flow (trapped chasers) ⇒ stronger reversal.

Data / spec. Top-10 crypto perps; 1h; 4h; 1,262 events (508 design / 754 test).

Primary result

Test CI [−0.0807, +0.0654]. Horizons: 1h −0.0067 · 4h −0.0022 · 24h +0.0068 · 3d +0.0063.

Artifacts: reports/cvd_failed_sweeps_study.md, reports/cvd_failed_sweeps_results.md, reports/cvd_failed_sweeps_findings.md.

Question. Do breakouts supported by persistent same-direction aggressive flow continue more than breakouts with weak flow or immediate flow reversal — after accounting for the breakout itself?

Signal. Breakout = close beyond the trailing 48-bar prior-range extreme (excludes current bar; 24-bar cooldown). Aligned, activity-normalised flow over 4 bars, z-scored (flow_align_z) plus signed persistence d·mean(sign(delta)). Outcome = oriented forward return, continuation-failure flag, MFE/MAE.

Prior-work overlap. breakout_regime_study.md already tested the single-bar filter cvd_20>0 (validate lift 1.07, not significant). H3 adds multi-bar persistence, displacement, failure-rate outcomes, a wider universe, and an explicit incremental-vs-price-only test.

Primary result — two universes

62-crypto horizons (test IC): 1h −0.0179 · 4h −0.0146 · 24h −0.0115 · 3d −0.0086 — negative everywhere.

Artifacts: reports/cvd_breakout_flow_study.md, reports/cvd_breakout_flow_results.md (+_all), reports/cvd_breakout_flow_findings.md.

• Everything failed. Across four independent order-flow hypotheses, every preregistered criterion failed, on both design and test.
• Recurring design→test sign flip. Divergence (+0.0248→−0.0040), H1 (+0.0049→−0.0187) and H2 (+0.0495→−0.0022) all looked least-bad in-sample and reversed out-of-sample — the classic selection artefact. H3 was negative in both windows.
• No stable parameter region in any of the robustness sweeps (18–27 cells each).
• Consistency with prior work. The existing breakout_regime_study already found the cvd_20>0 breakout filter insignificant; H3's result is directionally the same.
• Direction, where anything appears, is often the opposite of the hypothesis (H1 test, H3 intraday) — not merely zero.

Interpretation. At 1-minute→1-hour resolution, net taker flow (CVD) shows no incremental predictive value for future returns in this sample. The proxy nature of CVD (no maker intent, no hidden liquidity) plus the single-regime sample are the two standing limitations.

# 1 · Spot–perp divergence
python scripts/divergence_study.py

# 2 · H1 price absorption
python scripts/absorption_study.py

# 3 · H2 failed sweeps
python scripts/sweep_study.py

# 4 · H3 flow-confirmed breakouts
python scripts/breakout_study.py                    # 62 crypto perps
python scripts/breakout_study.py --include-noncrypto # 100 symbols

# full test suite (causality, alignment, synthetic edge/no-edge)
uv run pytest -q

Document & artifact map

All paths under reports/ unless noted. No prior reports were overwritten; no existing research code behaviour was changed. Reports were generated by the scripts above without inspecting results first (methods frozen in the design docs).