# Registered single-venue spot-flow experiment v1 — September 8, 2026

## Hypothesis and data

Test whether completed spot taker-buy imbalance improves an earlier-entry rule or invalidates a harmful short more effectively than the identical price condition alone. Source is official Binance BTCUSDT four-hour spot klines, not nioctib multi-exchange flow and not futures positioning. Taker-buy quote volume divided by total quote volume measures aggressor direction, not net new money entering Bitcoin. No expected profitable result or creator-accuracy claim is assumed.

Freeze monthly archives August 2017-August 2026 inclusive and daily September 1-7 2026 (116 archives), along with published SHA256 checksums, in a new data directory. Normalize documented pre-2025 millisecond and 2025-onward microsecond timestamps into integer Unix seconds. Validate twelve columns, exact four-hour close endpoints, ordering, period membership, positive valid OHLC, nonnegative finite volumes and taker buys no greater than total volume. Retain gaps and zero-volume bars as unavailable flow; no repair or interpolation. Record acquisition time and immutable file/normalized hashes. Complete the data receipt and commit its identity before model evaluation; no outcome inspection during intake.

Primary sources: https://github.com/binance/binance-public-data and https://developers.binance.com/docs/binance-spot-api-docs/rest-api/market-data-endpoints . Archives can be revised. Their historical bar timestamps do not prove original publication availability; modeled availability below is an assumption about a completed-bar collector, not proof these exact archived values were seen live. Binance USDT and Coinbase USD differ in venue/instrument; normalize flow as a ratio and retain Coinbase as the account price source.

## Flow feature and timing

For a decision at close time D, use exactly six contiguous four-hour flow bars whose latest close is D-L, where L is either four hours (primary) or 24 hours (availability stress). At each decision, buy share is sum(taker-buy quote volume) / sum(total quote volume) across those six bars. Strictly greater than 0.50 is positive buying. Every constituent bar must exist and have positive quote volume; otherwise feature is unavailable. No rounding before threshold comparison. The six-bar window is exactly 24 hours; gaps invalidate it rather than extending it backward. The lag is in addition to simulated trade execution delay. No use of a forming or same-decision flow bar. Retain aligned feature availability/values so signals can be audited.

## Fixed rules and paired controls

All price calculations use completed Coinbase four-hour bars with full prior price warmup; channels exclude the current candle and SMA includes it. Full price/flow state is formed before each account slice. No threshold search.

1. `price_short_veto`: start with original final targets. If target is negative, current close exceeds its 120-close SMA, and current close exceeds the close exactly six four-hour bars earlier, replace the short with cash for that decision. Otherwise keep the original target. Require all 120 closes and the exact 24-hour comparison timestamp.
2. `flow_short_veto_4h` and `flow_short_veto_24h`: same price condition, but replace the short with cash only when the corresponding available lagged buy share is strictly above 0.50. Missing flow means no override. Compare each directly with price_short_veto; comparison with original alone cannot establish added value from flow.
3. `price_early_entry`: independent long/cash state, flat until at least 120 preceding bars. Enter when price exceeds the prior 120-bar highest high, or when it exceeds the prior 30-bar highest high and its 120-close SMA. Exit on close below the prior 60-bar lowest low. Retain state otherwise. This adds a fixed five-day price shortcut to the original twenty-day breakout, retaining its ten-day exit. Exit has priority for a held long; flat entries occur on a later decision if applicable.
4. `flow_early_entry_4h` and `flow_early_entry_24h`: same state/exit/normal 120-bar entry; the 30-bar shortcut additionally requires corresponding available buy share above 0.50. Missing flow disables only the shortcut. A normal 120-bar breakout still enters; existing positions retain the same exit. Compare directly with price_early_entry as well as the unchanged breakout.

Three additional references: unchanged original final targets, unchanged 120/60 breakout, and BTC buy-once/hold. Total nine models: four flow variants, two paired price-only controls, three references. No leverage, feature training, future labels, momentum threshold tuning or adaptive candidate selection.

## Accounts and execution

Evaluate 14 windows: 2018-through-frozen-end; 2019-2022; 2023-through-frozen-end; annual 2018-2026 (partial 2026); 2019-through-frozen-end; and separate June 11-September 8 recent replay. The frozen end remains April 2026; recent prices use the separately pinned public cache. No pre-2018 return claim for this feature. Each window starts a new $10,000 account, with signals warmed on all preceding supplied history. No stitching winners across years.

Costs {6,30} bps per dollar traded and trade delays {1,3} four-hour bars: 9 x 14 x 4 = 504 cells. Primary 6 bps/4h execution delay. Four flow variants cover both feature-availability lags at each cost/delay. Reuse exact post-fee target accounting, fixed BTC quantity between fills, expiring scheduled fills at absent timestamps, and observed-close drawdown. No forced terminal sale; short funding/borrow, interest and intrabar liquidation remain omitted. These costs/delays are scenarios, not observed venue fills.

## Verification and interpretation

Verify pinned source/input and prior provenance identities. Reproduce 23,600 canonical decisions. Verify prior original/BTC/breakout curves for the 13 already evaluated matching windows, two costs and two delays (156 matching reference curves). The new 2018-through-end window has no previous reference curve. Verify future candle and future flow mutation cannot change prior features/signals/accounts. Explicitly test exact lag boundary, missing/zero-volume rejection, six-bar contiguity, normalization, unit transitions, price-only equivalence when all flows are positive, and original/breakout equivalence when flow is unavailable. Price-only controls must not change when flow changes.

Retain all 504 metrics and compressed traces [decision close timestamp, target, equity, BTC quantity]. Save aligned flow feature traces and coverage/missing reports separately. Report paired return/drawdown/cost/turnover/exposure differences by era, recent window and availability lag, including losses and hypotheses that fail. A flow candidate outperforming a weak original but trailing its price-only control does not demonstrate an informational improvement. Do not promote a historical winner, claim independent evidence from correlated cells, or infer a future advantage from this already explored history. New timestamped forward evidence and venue robustness remain required.