H-022: Forward replication of H-020 on data collected after 2026-10-15

— Family: flows · Status: registered

Do Bitcoin exchange inflows still predict drops on new data? Data is being collected. The pre-registered test runs on 2027-10-15, once a year of data has accumulated.

Forward data collected: 14 of 365 days (3.8%). First test on 2027-10-15.

Result

Sharpe: return per unit of risk (higher is better). CAGR: average yearly growth. Max drawdown: worst fall from a peak. p-value: chance the result is luck (lower is better). “Hold” columns are buy-and-hold for the same asset and days. Glossary

No engine results yet: this study is kept for the record and is not re-run.

The same rule on every coin

Exploratory check of this rule on all coins since 2022-01-01. It is not part of the verdict above.

✓works (passed the chance check)
↑better than holding, could be luck
✗worse than holding
·not enough data
original coin of the hypothesis

Coin	✓	Result	Sharpe: rule / holding	CAGR	From
BTC	↑	better than holding, could be luck	0.80 / 0.48	+29% / +12%	2022-01-01
ETH	↑	better than holding, could be luck	0.55 / 0.20	+17% / -9%	2022-01-01
SOL	↑	better than holding, could be luck	0.55 / 0.35	+14% / -10%	2022-01-01
BNB	↑	better than holding, could be luck	0.44 / 0.42	+10% / +8%	2022-01-01
XRP	↑	better than holding, could be luck	0.62 / 0.51	+23% / +10%	2022-01-01
ADA	↑	better than holding, could be luck	0.08 / -0.01	-18% / -32%	2022-01-01
LTC	↑	better than holding, could be luck	0.34 / 0.11	+1% / -18%	2022-01-01
LINK	↑	better than holding, could be luck	0.31 / 0.30	-5% / -10%	2022-01-01
DOGE	↑	better than holding, could be luck	0.52 / 0.26	+11% / -15%	2022-01-01
XLM	↑	better than holding, could be luck	0.47 / 0.30	+9% / -9%	2022-01-01
TRX	↑	better than holding, could be luck	0.80 / 0.74	+40% / +36%	2022-01-01
EVER	✗	worse than holding	-0.21 / -0.15	-65% / -68%	2022-03-19
GRAM	✗	worse than holding	-0.26 / 0.13	-32% / -17%	2023-03-14

Why it might work

H-020 passed on 2022+ data, but historical exchange-flow data can carry look-ahead bias: exchange addresses discovered later are labelled retroactively, so past flows look cleaner than they were in real time. This repeats the exact rule only on flows recorded day by day from 2026-10-15 on, as they were published. Tested after a year of such data.

Verdict history

Pre-registered specification

kind: timing_rule
rule:
  type: feature_vs_rolling_quantile
  feature: exnetflow_btc_7d
  q: 0.8
  long_when: not_above
  lookback: 365
  min_periods: 180
  lag: 1
mode: long_flat
fee: 0.001
universe:
- BTCUSDT
periods:
- name: forward
  start: '2026-10-15'
significance:
  method: circular_shift
  n: 2000
  alpha: 0.05
criterion:
  min_wins:
    forward: 1
  require_significance: true