H1 — Price Absorption: Findings & Recommendation

Status: complete — negative result.

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

Machine outputs: experiments/reports/absorption/{summary,by_symbol,leave_one_out,robustness}.json, panel.parquet.

Reproduce: python scripts/absorption_study.py.

1. Verdict

NO EDGE (H0). The preregistered criteria P1–P7 all fail. The impact residual (absorption) does not predict forward returns in the hypothesised direction; if anything it is mildly inverted out-of-sample and unstable.

2. What was tested

Rolling price-impact residual of aggressive perp taker flow, oriented forward returns, strong-flow events, matched controls, and a control regression — on the crypto-only top-10 perps, 1h bars, 4h primary horizon, 2026 (design Jan–Apr, test May–Sep). 1m klines only; no tick data.

3. Results

• Horizon IC (test): 1h −0.0171 · 4h −0.0187 · 24h +0.0014 · 3d −0.0035.
• Matched controls: absorbed − control = +17.2 bps, but CI [−7.6, +37.0] includes zero (n=3,657 events) → not significant, and inconsistent with the negative headline IC.
• Control regression: A coefficient ≈ 0.00001, CI [−0.00039, +0.00042] → indistinguishable from zero after controls.
• Per-symbol test IC: negative for 8 of 10 (HYPEUSDT +0.0203, BNBUSDT +0.0060 are the exceptions; OGNUSDT −0.0479, ETHUSDT −0.0379 weakest).
• Quintiles (strong-flow): non-monotone (−0.00018, −0.00079, −0.00057, +0.00008, +0.00049).

4. Robustness sweep (test IC)

18 cells (timeframe ∈ {15m,1h,4h} × window ∈ {15,30,60}d × event q ∈ {0.90,0.95}):

• Test IC is negative in every single cell (−0.0099 … −0.0233).
• Design IC is positive in most cells → a systematic design→test sign flip.
• Strong-flow-only IC is mixed and small.

This is the signature of a noise-driven in-sample artefact, not an edge.

5. Interpretation

The absorption story was mechanically plausible and the estimator was sound (a rolling-regression residual, not an unstable ratio). The data still reject it: any weak positive signal in the design window reverses out-of-sample and is indistinguishable from zero after controls and matching. Combined with the earlier spot–perp divergence negative, two independently motivated order-flow hypotheses now fail on the same 2026 sample.

6. Limitations (why the negative is not the final word)

• Single regime, in-sample 2026. Only BTCUSDT has multi-year history; the universe is 9 months. No true out-of-sample across regimes — the same limitation that sank the divergence study.
• Taker-flow proxy. CVD cannot see maker intent, hidden liquidity, or cancellations; absorption can be a venue artefact.
• n = 10 cross-section; 1m→1h resolution. Sub-minute impact dynamics are invisible without tick data.

7. Recommendation

1. Do not re-tune H1. Changing windows/quantiles to manufacture a positive result is not science; the sweep already shows the sign is stable-negative OOS.
2. Highest-leverage next step: build 2024–2025 perp 1m history (~2 h, ~3 GB; the downloader already supports it). This converts any flow hypothesis from exploratory to genuinely out-of-sample and should precede new order-flow bets.
3. Then test H3 (flow-confirmed breakout continuation) rather than H2. H3 has a clean price-only baseline and a natural tradeable framing, and does not depend on untested swing detection. H2 remains attractive if the interest is reversal structure, but it first requires building (and validating) causal swing-high/low detection, which does not exist in the repo.
4. H4/H5 as a market-state model, not a directional signal. Directional price-impact asymmetry (H5) may still be useful as an execution-cost / slippage model even where it has no alpha; keep it as infrastructure.
5. Do not acquire tick/aggTrades yet. The binding constraint is regime breadth and effect existence, not resolution. Reserve tick data for execution realism or if a 2024–2026 bar-level test of a specific hypothesis is promising.

Bottom line: abandon H1 as a directional signal; invest next in out-of-sample history (2024–2025), then H3.

← All reports · source (.md)