Do funding, OI, CVD and liquidation signals work?
Updated: 2026-10-27
We wrote down 6 rules built from the most-discussed crypto derivatives data before looking at results, then tested them on up to 2 years of Binance perpetual BTC, ETH and SOL. Short answer: none of them reliably beat random entries with the same direction and risk size.
Method
- Pre-registration: rule, thresholds and decision criteria were written before looking at the data and not changed afterwards.
- Trade: 1-hour bars; entry at the next bar's open, stop 1 × ATR(14), 1:1 target, at most 48 hours; if target and stop hit in the same bar → loss. No fees (gross results).
- Discovery and validation: each rule was chosen on one symbol and period, then tested unchanged on other periods and symbols.
- Comparison: for every real trade, a trade at a random hour in the same period with the same direction and risk size; 1,000 repetitions.
- Data: Binance perpetual public endpoints (candles, taker buy volume, funding, open interest); liquidations from a separate source.
Results
| Idea | Group | Trades | Win % | Avg R | p |
|---|---|---|---|---|---|
| CVD divergence | BTC year 1 (discovery) | 84 | 63% | +0.26 | 0.007 |
| BTC year 2 | 74 | 45% | −0.11 | 0.83 | |
| ETH+SOL year 1 | 212 | 49% | −0.02 | 0.61 | |
| ETH+SOL year 2 | 132 | 39% | −0.21 | 0.99 | |
| Against funding | BTC year 1 / 2 | 75 / 81 | 36% / 33% | −0.28 / −0.33 | 0.99 / 1.00 |
| Pooled validation | 406 | 45% | −0.10 | 0.96 | |
| With funding | BTC 2022–24 (unseen) | 165 | 55% | +0.10 | 0.057 |
| ETH+SOL 2022–24 | 364 | 44% | −0.12 | 0.98 | |
| Against liquidation spikes | BTC (30 days) | 25 | 44% | −0.12 | 0.73 |
| ETH+SOL (30 days) | 51 | 45% | −0.10 | 0.74 | |
| BOS/CHoCH (1H, ATR stop) | BTC (30 days, discovery) | 33 | 67% | +0.33 | 0.04 |
| ETH / SOL (30 days) | 32 / 41 | 50% / 42% | 0.00 / −0.17 | 0.57 / 0.89 |
How to read: avg R is the mean result per trade in risk units (+1 = target, −1 = stop). p is the share of 1,000 random runs whose average matched or beat the rule. The OI-unwind rule produced almost no signals in 30 days of OI history and could not be tested.
What we learned
The most instructive result is CVD divergence: on one year of BTC it looked like a ‘working signal’ (63% hit rate, p = 0.007); the same rule lost money the next year and on ETH/SOL. Even a very strong result on one symbol and one period can be luck. Trading against funding was negative everywhere; trading with funding was borderline on BTC (p = 0.057, gross) and did not hold on ETH/SOL.
That is why ByStock shows these data as market context, not as trading signals: which side is crowded, whether positions are opening or closing, whether aggressive orders agree with price. The ‘market context’ card on each crypto outlook page explains them in plain language.
Limitations
- Only 3 symbols and 1-hour bars were tested; other timeframes or exit rules may differ.
- Liquidation, OI and BOS/CHoCH tests used only 30 days of data; samples are small.
- Results exclude fees (gross); real costs would make them worse.
- Past results do not guarantee the future.
Frequently asked questions
Does shorting when funding is high work?
Not in our test. A rule that shorts when the 7-day funding z-score on Binance perpetuals is above +2 and goes long below −2 (1-hour bars, ATR stop, 1:1 target) lost money on BTC in two separate years and did worse than random entries over 406 validation trades including ETH and SOL.
Is CVD divergence a reliable signal?
It looked excellent in one year of BTC (63% hit rate, p = 0.007), but the same rule lost money in the following BTC year and in two years of ETH and SOL. A strong result on one symbol and one period can still be luck.
So are these data useless?
In the forms we tested as trading signals, they showed no edge. They are still useful for understanding the market (which side is crowded, whether positions are opening or closing), which is why ByStock shows them as context, not signals.
What does comparing with random entries mean?
For every real trade, a trade is opened at a random hour with the same direction and the same risk size; this is repeated 1,000 times. p is the share of random runs whose average result matched or beat the rule. A small p means the rule beat random.