Williams %R + momentum on BTC 5-minute markets
A viral post credits this setup for about $1,364 a day on Polymarket's BTC "Up or Down" markets. We tested the strategy as described.
Verdict: no edge. As described, the strategy is a coin flip (49–50% correct) and loses money after costs in both periods we tested. It never beats picking random minutes.
The claim
The post describes a reversal system: Williams %R finds when BTC is at an extreme of its recent range, and momentum confirms the turn. Bet Up when %R is below −80 and momentum turns positive; bet Down when %R is above −20 and momentum turns negative.
What we tested
- Williams %R (14) and momentum (10) on BTC 1-minute candles (Bybit).
- Outcome: is BTC higher or lower 5 minutes after the signal (Polymarket's up/down rule; a flat close counts as Up).
- Cost: 1 cent per $1 bet, so break-even is 50.5% correct.
- Two separate periods: Oct 2019 – Dec 2022 and Jan 2023 – Oct 2025.
- Compared with entering at random minutes.
- The rules were written down and frozen (SHA-256 fingerprint 25647940…) before any test ran. Nothing was tuned afterwards.
Results
| 2019–22 signals | % correct | 2023–25 signals | % correct | |
|---|---|---|---|---|
| All | 9,254 | 50.3% | 6,721 | 49.3% |
| Up bets | 4,353 | 52.8% | 3,196 | 50.8% |
| Down bets | 4,901 | 48.1% | 3,525 | 47.8% |
- After costs: −0.4 cents per $1 in 2019–22 and −2.5 cents per $1 in 2023–25.
- Against random entry: random minutes score about 50% too. The signal ranks at the 79th and 13th percentile of random runs; beating chance would need the 95th.
- The Up bets look better than they are: because a flat close counts as Up, a random Up bet already wins 51.5%.
- Where good-looking screenshots come from: a few lucky weeks, such as 75.9% on 29 bets (Aug 2022). Only about 3 of 150 weeks in each period clear 65%.
What this does not show
This tests the strategy as described in the post, not the account's actual trading. A bot can make money in these markets in other ways, for example by pricing faster than the market. It also says nothing about data after 2 October 2025, which we left untouched.
Full evidence
- Frozen rules: the exact pre-registered test and its fingerprint.
- Code: the script that ran the test, to re-run or adapt.
- Every signal: all 15,975 signals with time, direction and outcome.
- Full statistics: random-entry comparison, results by hour and side, 10- and 15-minute horizons.
- Data details: source, dates and gaps.
Available on zx10 when it opens.