Absorption: heavy taker flow with a flat price does not predict a reversal
If market buys arrive in a wave and price does not move, a large passive seller is often assumed to be absorbing them, so price should then fall (and the mirror for sells). We tested this on 40 Binance perpetuals over six and a half months and on 30 instruments across three venues in ArcDelta. A flat price adds nothing: after a large dollar flow price leans against it slightly over 4 hours whether or not it moved, and the effect does not cover taker costs.
| Data and condition | Holdout events | P(against flow) vs base | Edge | After taker costs | Verdict |
|---|---|---|---|---|---|
| ArcDelta · imbalance z ≥ 2, |r| ≤ 0.5σ · 15m | 512 | 47.7% / 47.8% | −0.2 pp | −16.8 bp | Unsupported |
| Binance · CVD z ≥ 2.5, |r| ≤ 0.5σ · 4h | 1 123 | 51.9% / 49.6% | +2.3 pp | −6.7 bp | Inconclusive |
| Binance · CVD z ≥ 2.5, any move · 4h | 3 906 | 52.9% / 49.8% | +3.2 pp | −6.6 bp | Control |
| Binance · CVD z ≥ 2.5, |r| > 1σ · 4h | 3 604 | 52.6% / 49.8% | +2.8 pp | −9.4 bp | Control |
The claim
Condition: the 15-minute taker flow is at least 2–3 standard deviations from the instrument's last three days, while the 15-minute return stays within 0.25 or 0.5 of its usual size; optionally open interest rose. Outcome: the return over 15 minutes, 1 hour or 4 hours goes against the flow. Flow was measured two ways, as dollar CVD and as the buy–sell imbalance share of turnover.
How it was tested
72 variants per data set (flow measure, threshold, flatness, open-interest filter, horizon), all corrected together with Benjamini–Hochberg at q = 0.1. Events start when the condition turns true and are spaced by the horizon on each instrument. The base rate is the frequency of the same outcome on the same instrument at the same hour of day. P-values account for events that fire together across instruments. The trade result books a taker round trip of 14 bp. To count, a variant needs at least 5 percentage points over the base on the holdout, the lower end of its 95% interval above the base, at least 100 events, the same sign in both halves and on most instruments, and a positive result after costs.
What we found
On three weeks of ArcDelta data no variant survived the correction. On six months of Binance data eight variants did, all at 4 hours and all on dollar CVD, but on the holdout their edge was 1.4–3.2 points, below the 5-point threshold, and the average trade lost money after costs. The control rows show why this is not absorption: the same edge appears when price had already moved strongly in the direction of the flow. What remains is a weak, general lean against very large flow over a few hours, too small to trade on taker fees.
Limits
Binance is one venue; the multi-venue ArcDelta sample covers only three weeks. Order-book walls on the absorbing side were not tested because minute book history is still short. Five-minute bars blur sub-minute absorption, which a second-level archive will allow us to test later.
Period & evidence limits
Binance USDT-M 5m bars with taker buy volume, Apr 1 – Oct 18, 2026 (discovery before Sep 1, holdout from Sep 1); ArcDelta 5m bars with taker flow summed over Binance, Bybit and OKX, Oct 1–19, 2026 (holdout from Oct 13). Results require sample support and independent evaluation. Publications exclude private parameters, source code and user data.