Predicting funding rates before the market moves

Perpetual futures never expire. That's their whole thing. But they still need to stay pegged to spot price somehow. The mechanism for this is funding rate, a periodic payment between longs and shorts.

Every 8 hours, one side pays the other:

• Positive funding: longs pay shorts
• Negative funding: shorts pay longs

If you know where funding is going before it gets there, you can position accordingly. That was the goal.

What I tried

Started with the obvious stuff. Historical funding rates going back 30 days, open interest changes, order book imbalances. The features looked like this:

features = [
    'funding_rate_lag_1',
    'funding_rate_lag_2', 
    'oi_change_pct',
    'bid_ask_imbalance',
    'volume_delta'
]

Open interest changes turned out to be surprisingly useful. When OI spikes without price movement, someone is positioning. The order book data was noisier than I expected.

Normalizing the signal

Raw funding rates are noisy. Small movements happen all the time and mean nothing. I used z-scores to filter for statistically significant moves:

z = (x - mean) / std

Anything above 2 or below -2 gets flagged as a potential signal. Everything else is ignored.

What actually worked

After 12 months of backtesting on BTC perpetuals:

72% directional accuracy sounds good until you realize the edge is thin. Transaction costs eat into it. Slippage eats into it. Still, it works well enough to be part of a larger system.

The model lives in Calibrasteme now, running alongside the other signals.