Generating Alpha from Funding-Rate Basis on Hyperliquid

Hyperliquid's on-chain perpetual engine publishes a predicted funding rate every hour, settled every 8 hours. That prediction is public, deterministic, and — critically — often misprices the actual cost of carrying a synthetic long versus holding spot. When it does, the spread is a carry signal. When it does not, chasing it is how you donate to informed flow. What follows is the exact framework we run in production to separate one from the other.

The Basis Arithmetic

The spot-perp basis for asset $X$ at time $t$ is:

basis(t) = (perp_mid(t) - spot_mid(t)) / spot_mid(t)

Annualised, this should equal the funding rate over the same horizon. Hyperliquid reports predicted_funding as an 8-hour rate; to annualise: multiply by 3 (three 8-hour windows per day) and by 365. The carry signal is then:

carry(t) = annualised_predicted_funding(t) - annualised_basis(t)

A positive carry(t) means the market is pricing in more funding income than the basis alone justifies — the perp is rich relative to spot. You short the perp and buy spot, collecting the spread. A negative value means the opposite and the position flips. The edge lives in the magnitude: entries below roughly 4% annualised after costs are noise.

Transaction-Cost Reality

The numbers that kill funding strategies on paper are the ones that actually kill them live. On Hyperliquid the taker fee is 2.5 bps; maker rebate is 0.2 bps. Spot legs — whether routed through Jupiter aggregator or a direct DEX pool — run 1–3 bps depending on liquidity depth and slippage. Add funding settlement variance (actual rarely equals predicted by more than 15%, but tail deviations reach 40% in distress) and you need the raw carry signal to clear at least 12–15 bps per 8-hour window before entering.

We model round-trip cost as:

total_cost = 2 * (taker_perp + spot_slippage_estimate) + funding_variance_reserve

In backtests on BTC and SOL from January 2024 through March 2025 using Hyperliquid's historical tick data, entries that cleared the 12-bps threshold captured 71% of available carry days while cutting loss-making entries from 38% to 11%. That 27-percentage-point improvement in win rate entirely justified being out of the position 29% of the time.

Regime Filter: Don't Harvest Into a Storm

The single biggest source of adverse selection in funding strategies is entering a carry trade right before a sharp directional move. Funding rates spike precisely when conviction is high and positioning is extreme — the same conditions under which a mean-reversion in funding can coincide with a punishing spot move in the delta leg.

The filter we use is a realised-volatility gate. We compute 4-hour realised vol on spot prices (5-minute returns, 48 observations), annualise, and compare to a trailing 30-day 75th percentile. If current vol exceeds that threshold, we treat the signal as contaminated and either close existing positions or skip new entries entirely.

In practice this means:

• Below threshold: enter when carry signal clears cost hurdle, maintain position through settlement.
• Above threshold: hold cash, accept zero carry for the window rather than risk delta bleed eating multiple windows of harvested funding.
• At entry, always check: is the perp order book depth within 0.5% of mid above $500k? Thin books mean your entry itself moves the basis against you.

The vol filter removed approximately 18% of would-be entry windows in backtesting and eliminated every instance of a loss exceeding 3x the expected carry gain for that window. That asymmetry is the entire point.

Position Sizing and Rebalancing

Delta neutrality requires continuous attention. If spot moves 2% intraday, the delta on a short-perp / long-spot book drifts. We rebalance when the gross delta imbalance exceeds 0.3% of notional — small enough to stay neutral, large enough to avoid over-trading.

Position sizing is Kelly-fractional, capped at 15% of capital per single asset. In production we run three to five uncorrelated assets simultaneously; the aggregate carry income has a Sharpe north of 2.1 over the backtest window, driven mostly by the regime filter and the cost discipline rather than any exotic signal engineering.

The automated execution layer handles rebalancing, settlement tracking, and vol-filter computation entirely on-chain where possible, falling back to off-chain WebSocket feeds for latency-sensitive order submission.

What the Backtest Does Not Capture

Be honest with yourself about backtest limitations here. Three things the numbers do not fully price in:

1. Liquidity regime change. Hyperliquid's open interest has grown 6x since early 2024. Slippage estimates from thin early markets understate current costs in some assets and overstate them in majors.
2. Oracle manipulation risk. Funding on Hyperliquid references an index; a spoofed index spike can cause a transient funding rate that looks harvestable but resolves against you before settlement.
3. Smart contract upgrade risk. The protocol has shipped breaking changes to funding mechanics twice in the past 18 months. Any production system needs to re-validate signal assumptions after each upgrade.

None of these are dealbreakers, but they mean the edge is real and fragile, not real and permanent. We monitor for oracle divergence (>50 bps between Hyperliquid index and CEX reference price) and pause automatically when it fires.

---

If you want to run a version of this strategy in production rather than in a notebook, reach out — we scope, build, and operate the full execution stack so you can focus on signal research, not infrastructure.