Out-of-Sample Order-Flow Hypotheses — design 2025 / test 2026

• design: 2025-01-01 → 2026-01-01 · test: 2026-01-01 → 2026-10-01
• universe: 58 crypto perps (51 cross-market for divergence)
• These are the frozen H-series methods re-run on a genuine temporal split; verdicts are provisional (2025 vs 2026 is still only two regimes).

Spot–Perp CVD Divergence — Results (exploratory, in-sample 2026)

• universe (51): AAVEUSDT, ADAUSDT, ALGOUSDT, APTUSDT, ARBUSDT, ASTERUSDT, AVAXUSDT, BCHUSDT, BNBUSDT, BTCUSDT, CRVUSDT, CTSIUSDT, DOGEUSDT, DOTUSDT, ENAUSDT, ETHFIUSDT, ETHUSDT, FETUSDT, FILUSDT, GTCUSDT, HBARUSDT, INJUSDT, JUPUSDT, LINKUSDT, LITUSDT, LTCUSDT, METUSDT, MINAUSDT, NEARUSDT, OGNUSDT, ONDOUSDT, ORCAUSDT, PENGUUSDT, PUMPUSDT, PYTHUSDT, QNTUSDT, RLCUSDT, SANDUSDT, SOLUSDT, STRKUSDT, SUIUSDT, TAOUSDT, TIAUSDT, TRUMPUSDT, UNIUSDT, WLDUSDT, WUSDT, XLMUSDT, XRPUSDT, ZECUSDT, ZROUSDT
• timeframe: 1h · primary horizon: 4h
• window=60 · norm_window=720

Notes

• Exploratory only: single regime (2026), in-sample, n=51 cross-section.

Primary: pooled rank IC

test IC 95% bootstrap CI: [-0.0091, 0.0120]

Verdict: NO EDGE (H0)

Test deciles (D → forward return, market-neutral)

Test event study (|D| quintiles)

Test leave-one-symbol-out pooled IC

H1 — Price Absorption — Results (exploratory, 2026 in-sample)

• universe (58): 1000PEPEUSDT, AAVEUSDT, ADAUSDT, AKEUSDT, ALGOUSDT, APTUSDT, ARBUSDT, ASTERUSDT, AVAXUSDT, BCHUSDT, BNBUSDT, BTCUSDT, CRVUSDT, CTSIUSDT, DOGEUSDT, DOTUSDT, ENAUSDT, ETHFIUSDT, ETHUSDT, FETUSDT, FILUSDT, GTCUSDT, HBARUSDT, HYPEUSDT, INJUSDT, JCTUSDT, JUPUSDT, LINKUSDT, LITUSDT, LTCUSDT, METUSDT, MINAUSDT, NEARUSDT, OGNUSDT, ONDOUSDT, ONUSDT, ORCAUSDT, PENGUUSDT, PUMPUSDT, PYTHUSDT, QNTUSDT, RAYSOLUSDT, RLCUSDT, SANDUSDT, SOLUSDT, STRKUSDT, SUIUSDT, TAOUSDT, TIAUSDT, TRUMPUSDT, UNIUSDT, VVVUSDT, WLDUSDT, WUSDT, XLMUSDT, XRPUSDT, ZECUSDT, ZROUSDT
• timeframe 1h · horizon 4h · impact_window 720 · event q0.90

Notes

• Exploratory: single regime (2026), in-sample; n=58 perp symbols.

Primary: pooled rank IC(A, oriented forward return)

test IC 95% bootstrap CI: [-0.0206, 0.0014] strong-flow-only test IC: -0.0161

Horizon consistency (test IC)

Test quintiles (A → oriented return, strong-flow)

Matched control (absorbed − control, strong-flow)

• spread: 5.78 bps (CI [-5.37, 17.21])
• n events: 38254

Control regression (A coefficient)

• coef: 0.00002 (CI [-0.00024, 0.00033])

Per-symbol test IC

Robustness sweep (test IC; sign/magnitude stability)

Verdict: NO EDGE (H0)

H2 — Failed Liquidity Sweeps — Results (exploratory, 2026 in-sample)

• universe (58): 1000PEPEUSDT, AAVEUSDT, ADAUSDT, AKEUSDT, ALGOUSDT, APTUSDT, ARBUSDT, ASTERUSDT, AVAXUSDT, BCHUSDT, BNBUSDT, BTCUSDT, CRVUSDT, CTSIUSDT, DOGEUSDT, DOTUSDT, ENAUSDT, ETHFIUSDT, ETHUSDT, FETUSDT, FILUSDT, GTCUSDT, HBARUSDT, HYPEUSDT, INJUSDT, JCTUSDT, JUPUSDT, LINKUSDT, LITUSDT, LTCUSDT, METUSDT, MINAUSDT, NEARUSDT, OGNUSDT, ONDOUSDT, ONUSDT, ORCAUSDT, PENGUUSDT, PUMPUSDT, PYTHUSDT, QNTUSDT, RAYSOLUSDT, RLCUSDT, SANDUSDT, SOLUSDT, STRKUSDT, SUIUSDT, TAOUSDT, TIAUSDT, TRUMPUSDT, UNIUSDT, VVVUSDT, WLDUSDT, WUSDT, XLMUSDT, XRPUSDT, ZECUSDT, ZROUSDT
• timeframe 1h · horizon 4h · lookback 48 · cooldown 24
• sweep events (design+test): 17109

Notes

• Exploratory: single regime (2026), in-sample; n=58 perp symbols.

Primary: event pooled rank IC(af_z, oriented return)

test IC 95% bootstrap CI: [-0.0201, 0.0354]

Horizon consistency (test event IC)

Test quintiles (af_z → oriented return)

Matched control (high-af − low-af sweeps)

• spread: -0.81 bps (CI [-23.10, 23.49])
• n: 7756

Control regression (af_z coefficient)

• coef: 0.00020 (CI [-0.00031, 0.00081])

Per-symbol test event IC

Robustness sweep (test event IC)

Verdict: NO EDGE (H0)

H3 — Flow-Confirmed Breakout Continuation — Results (exploratory, 2026)

• universe (58): 1000PEPEUSDT, AAVEUSDT, ADAUSDT, AKEUSDT, ALGOUSDT, APTUSDT, ARBUSDT, ASTERUSDT, AVAXUSDT, BCHUSDT, BNBUSDT, BTCUSDT ...
• timeframe 1h · horizon 4h · lookback 48 · flow_window 4
• breakout events (design+test): 15467

Notes

• Exploratory: single regime (2026), in-sample; n=58 perp symbols.
• Design IC positive but flipped sign out-of-sample — selection artifact, not an edge.

Primary: event pooled rank IC(flow_align_z, oriented return)

test IC 95% bootstrap CI: [-0.0551, 0.0105]

Horizon consistency (test event IC)

Test quintiles (flow_align_z → return / failure rate)

Matched control (high − low alignment)

• spread: 11.16 bps (CI [-17.08, 38.67])
• n: 6949

Incremental regression (flow_align_z coefficient)

• coef: -0.00124 (CI [-0.00250, -0.00006])

Per-symbol test event IC

Robustness sweep (test event IC)

Verdict: NO EDGE (H0)

← All reports · source (.md)