Graveyard / Positioning / Binance perps
Open interest trading strategies
Does a rise or fall in Bitcoin open interest predict price?
The idea
Open interest is the total number of open futures contracts. A sharp rise can mean leverage is building, a sharp fall can mean liquidations. Traders use it to follow breakouts or to fade crowded positioning.
How we tested it
| Data | BTCUSDT perpetual futures (and ETH for one pooled test). Signals: OI percent-change z-score with direction confirmation; the same faded; OI drop with price divergence; Binance minus Bybit OI divergence; OI expansion as a gate on breakouts; funding z-score plus OI growth squeeze fade; funding z-score above 3 sigma with OI and taker flow. |
|---|---|
| Period | About two years of open-interest history (the free Binance endpoint only goes back about 30 days, so a cached two-year file was used). |
| Costs | Not stated for the individual tests. |
| Holdout | Full period, in-sample and out-of-sample gates with a minimum of 50 trades. |
| Control | None stated. |
Result
| Signals tested | 9 variants, both directions |
|---|---|
| OI momentum | fell short on 3 of 6 gates, gains in one bull segment |
| OI fade | loses at short and medium windows |
| Funding plus OI squeeze (BTC) | 0 of 36 configurations reached 50 trades |
| Funding plus OI squeeze (BTC and ETH pooled) | 0 of 36, 97 baseline events |
| Best small-sample configuration | Sharpe 0.67, profit factor 2.04 on 3 to 12 trades |
| Binance vs Bybit OI divergence correlation | 0.008 to 0.053 against a bar of 0.13 to 0.21 |
| Breakouts confirmed by rising OI | worse than unconfirmed ones |
Why it failed
Open interest in both directions fails. Momentum is a regime bet, the fade loses, and the confirming filter works backwards. Combined filters fire too rarely to test: a rule that needs several conditions at once produced 1 to 47 trades per configuration. One early version used a threshold 14 times the real maximum, so it never fired. The cross-exchange version also hits a data wall because the free endpoints keep only about 30 days.
What would make it worth retesting
A different open-interest source with several years of real depth, a looser composite score instead of strict AND-gates, or pooling more symbols. The notes treat the family as exhausted until then.
Past results of a backtest say nothing certain about the future.