H-026: Short-term reversal: hold the 3 coins that fell the most yesterday

Fail Family: other · Status: rejected

Does buying yesterday's biggest crypto losers pay? The idea did not beat simply buying the coin and holding it. It is rejected.

Result

since_2020: does not beat the equal-weight basket; since_2022: does not beat the equal-weight basket

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

Asset	Period	Sharpe	Hold Sharpe	CAGR	Hold CAGR	Max drawdown	Hold drawdown	In market	p-value	Beats hold	Counts as win
PORTFOLIO	since_2020 (2020-01-01 → 2026-10-20)	0.17	1.13	-22.3%	+75.6%	-99.3%	-77.0%	100%	0.908	no	no
PORTFOLIO	since_2022 (2022-01-01 → 2026-10-20)	-0.82	0.41	-54.0%	+6.5%	-98.1%	-67.0%	100%	0.990	no	no

Equity curves

Growth of 1 unit, log scale, fees included. Only days on which the rule is fully defined are counted.

Asset and period:

Why it might work

Proposed by ChatGPT (via F1DevBin) as its second experiment: after a sharp move relative to the other coins, part of it is given back over the next hours. Daily version (the proposal uses 1-6 hours). Known risk, also noted in the proposal: the effect may exist but be eaten by fees, since the basket changes almost every day.

Verdict history

2026-10-21: Fail (this engine, data to 2026-10-20)

Pre-registered specification

kind: cross_section
signal:
  type: reversal
  lookbacks:
  - 1
select:
  top: 3
fee: 0.001
universe:
- BTCUSDT
- ETHUSDT
- SOLUSDT
- BNBUSDT
- XRPUSDT
- ADAUSDT
- LTCUSDT
- LINKUSDT
- DOGEUSDT
- XLMUSDT
- TRXUSDT
periods:
- name: since_2020
  start: '2020-01-01'
- name: since_2022
  start: '2022-01-01'
significance:
  method: circular_shift
  n: 1000
  alpha: 0.0167
criterion:
  min_wins:
    since_2020: 1
    since_2022: 1
  require_significance: true