# Volume spike without a price move — frozen pilot protocol

Frozen before first outcome calculation on 2026-09-25 UTC. Main question: after a high-volume but narrow 1-hour Binance spot candle, do the next 1, 4 and 24 hours show follow-through or reversal, and does contemporaneous valid taker-buy share distinguish paths? This is a **six-major-pair observational pilot**, not evidence of accumulation or distribution by any named actor.

## Universe and timing

BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, XRPUSDT and DOGEUSDT; common UTC interval `[2025-09-25T06:00:00Z, 2026-09-25T06:00:00Z)`. Binance official/reconstructed hourly candles archived retrospectively; require contiguous bars. Event features at completed candle T use no future values. The first 24 hours warm up the baseline. Future returns use the close of T+1, T+4, T+24 relative to T close; incomplete horizons are pending/missing, never zero.

## Event and comparators

- Flat candle: `abs(close/open-1) <= 0.0025` (0.25%). High-volume candidate: quote volume >=2 × median quote volume of the **preceding** 24 completed 1h bars, not including T. Both conditions required; zero/invalid baseline excluded.
- Valid active-buy share: `taker_buy_valid=1`, `qv>0`, `0<=taker_buy_qv<=qv`. The event can still be counted if flow is unavailable, but flow subgroup is labeled missing. Valid groups: sell-leaning <45%; balanced 45–55% inclusive; buy-leaning >55%.
- Retain the first candidate, then a 24-hour per-symbol cooldown after the previous retained event. Publish excluded duplicate candidates. Non-spike flat hours (`relative_volume<2`) with complete windows form a **descriptive**, unmatched comparator; they are not a causal counterfactual and can overlap other events.
- Main result is median forward return and positive fraction at 1/4/24h, with 24h MFE/MAE from subsequent high/low. Show per-symbol and flow subgroups, not only pooled numbers.
- A 1,000-draw UTC-week cluster bootstrap of the median 24h event return (seed 250925) is a dependence-aware descriptive interval, not a calibrated probability of future profit. No best threshold is selected after results. Primary sample minimum: 20 retained, complete 24h events; below this, publish exploratory/insufficient only.

## Public evidence and limits

Release frozen input bars (`inputs.csv.gz`), every evaluated hourly candidate (`candidate-ledger.csv.gz`), retained events (`events.csv`), comparator rows (`controls.csv.gz`), summaries and bootstrap seed/contract, SHA-256 manifest, and an independent verifier. Chart is generated from the exact release. Historical reconstruction does not prove information was available locally at the event time. One-hour OHLC and taker-buy fields do not reveal passive absorption, wallets, hidden orders, order-book cancelations or who traded. Spot buying may coexist with price weakness; do not call the pattern a proven buy signal.