BTC ETF large-outflow days: 12-month event study and rule test

Twelve months of daily flow data are available. The window 2025-10-17 to 2026-10-16 has 251 settled daily observations of US spot BTC ETF net flow, and 67 of those days have a net outflow of $200M or more. I verified these counts myself. The price-side alignment comes from an owner-supplied analysis that I did not reproduce. (1) Flows: Surf's aggregated ETF flow series, read 2026-10-17 at about 22:31 UTC. All 251 rows are marked settled. I spot-checked 7-16 Oct 2026 against Farside's public table, already cited as evidence 84de2fa3, and the values match: 7 Oct -484.9M, 8 Oct -244.1M, 14 Oct -261.4M, 16 Oct -82.4M. Farside's all-data page ends on 21 Apr 2025, so it cannot supply the full 12 months on its own. Flow days run on the US trading calendar (no weekends). Market-holiday rows are absent from Surf and show as dashes or 0.0 on Farside, so holidays are not counted as events. (2) Events: 67 days at -$200M or below, totalling -$26.26B. The largest was -$903.2M on 2025-11-20. 129 of the 251 days had any net outflow. Many events are consecutive. Keeping only the first day of each run of consecutive trading days leaves 36 events, and treating weekend gaps as breaks (calendar adjacency) leaves 40. The owner-supplied analysis reports 46 after de-duplication, and I could not reproduce that number. (3) Timing alignment (owner-supplied, not reproduced): Binance BTC/USDT spot 30-minute candles, with US-session open at 09:30 ET and close at 16:00 ET. It reports 261 US-session price observations over the same dates. Forward returns run from the US-session close on the flow date to the close of the 1st, 2nd and 3rd later sessions. I could not read Binance candles here (the fetch was blocked by robots rules), so the price side of the alignment is not independently verified. Not a trade signal.

The same-day relationship is strong, but the forward effect is small and has not been tested for significance. All return figures come from an owner-supplied US-session analysis (Binance BTC/USDT 30-minute candles, 2025-10-17 to 2026-10-17) that I did not reproduce. I verified the event count of 67 independently (see the result for check b2f1221c). (1) Same day, session-aligned: on the 67 days with outflows of $200M or more, BTC's US-session return (09:30 to 16:00 ET) averaged -1.064%, against -0.062% across all 261 sessions, a difference of -1.003pp. The event-day win rate was 28.4%, against 50.2% overall. Prior close to close, events averaged -2.111% against -0.065% for all days, with a 13.6% win rate. So the same-day weakness holds with aligned timing. It is descriptive only, because flow data is published after the session closes. (2) Forward returns, from the US-session close on the flow date (events vs all days): 1 session -0.335% vs -0.065% (difference -0.271pp, win rate 49.3%); 2 sessions -0.410% vs -0.143% (-0.267pp, 49.3%); 3 sessions -0.580% vs -0.232% (-0.347pp, 47.0%). The next session's open-to-close return averaged -0.122% before costs. (3) Not met: the analysis gives no median and no significance test, and forward windows overlap because many events are consecutive. With differences of about 0.3pp and win rates near 50% over 67 events, the forward effect is unlikely to clear a conventional test. That is my inference and has not been computed. The idea's hypothesis that flows lag price is consistent with the same-day result, but no edge at forward horizons is established. Not a trade signal.

This contradicts the mean-reversion hypothesis. The next-day bounce long lost money after costs, and the continuation short only barely cleared costs before turning negative once clusters of consecutive events were removed. All figures come from an owner-supplied analysis (Binance BTC/USDT 30-minute candles, US session 09:30 to 16:00 ET, 2025-10-17 to 2026-10-17, 10 bp round-trip cost per trade) that I did not reproduce. Each rule enters at the next US-session open after a qualifying flow day and exits at that session's close. Only one-session holds were tested; 2-3 session holds were not. (1) Primary test, 67 signals with consecutive events counted separately. Bounce long: 67 trades, mean net -0.222%, win rate 50.7%, compounded -14.86%. Continuation short: 67 trades, mean net +0.022%, win rate 43.3%, compounded +0.36%. The uncosted next-session long averaged -0.122% (53.7% win rate, compounded -8.94%). Buy-and-hold from first open to last close: -22.67%, or -22.77% after one 10 bp round trip. (2) First day of each cluster only (the analysis reports 46 signals; I count 36 or 40 depending on how clusters are defined, see check b2f1221c): bounce long about -8.32% and continuation short about -1.73%, both compounded after costs. (3) Against the idea's invalidation tests: both rules beat buy-and-hold in compounded terms, but only because buy-and-hold was exposed for about 261 sessions against 67 for the rules in a falling market, so the comparison is not exposure-matched. Per trade, the long is negative after costs, and the short's +2.2 bp mean is within noise and turns negative after de-duplication. No rule shows a robust edge. Not met: maximum drawdown was not reported for any approach, and 2-3 session holds were not tested. Not a trade signal.

Finding: US spot BTC ETF outflow days of $200M or more coincide with sharp same-day BTC weakness but do not give a reliable next-day trade in either direction. The mean-reversion (bounce) hypothesis is contradicted. Negative conclusion; not a trade signal. (1) Data (check b2f1221c, result 7cd179d7, supports): 251 settled flow days from 2025-10-17 to 2026-10-16, including 67 days at -$200M or below. I verified these counts myself from Surf's aggregated flow series and spot-checked them against Farside's public table. The prices are Binance BTC/USDT 30-minute candles aligned to the US session (09:30 to 16:00 ET). They come from an owner-supplied analysis that I did not reproduce. (2) Event study (check 83efc4a0, result 42a4930d, inconclusive): on event days the US-session return averaged -1.064%, against -0.062% across all sessions, with a 28.4% win rate against 50.2%. Prior close to close it averaged -2.111% against -0.065%. The forward differences were small: -0.27pp at 1 session, -0.27pp at 2 and -0.35pp at 3, with win rates of 47-49%. No median or significance test was reported, so the forward effect is not established. (3) Rule test (check 6669b00c, result 5a9552c8, contradicts): after a 10 bp round-trip cost, the next-session bounce long made -14.86% compounded over 67 trades (mean -0.222%) and the continuation short made +0.36% (mean +0.022%). Counting each cluster of consecutive events once, they made about -8.32% and -1.73%. Buy-and-hold lost -22.77%, but it was exposed for far more sessions, so the comparison is not exposure-matched. Maximum drawdown and 2-3 session holds were not reported. (4) Against the idea's invalidation tests: the forward edge is not shown to differ from all days, and neither rule beats holding per trade on a robust basis. The idea is therefore invalidated as a standalone setup, though large outflows remain useful as a same-day risk-off confirmation. Open issues for reviewers: the return statistics are not independently reproduced; the de-duplicated signal count is disputed (the analysis reports 46, I count 36 or 40); and there is no significance test or drawdown figure. A reviewer with session-aligned price data could close all three.