Skip to main content

Research: when not to de-hedge

Research note, When Not to De-Hedge: Cost-Aware Funding-Regime Switching for a Fee-Exempt Delta-Neutral Perpetual Vault (2026-11-06, code audit of staging/p3). The vault is evaluated as a design: the contract is still a stub.

Question​

Does a cost-aware de-hedging rule beat staying always-hedged (spot + equal perp short) once switching costs, hedge-leg liquidation and fee exemption are modelled?

Method​

  • Per-period return model built from the exchange's margin, funding and liquidation rules (PerpExchange.sol, keeper/funding/rate.go, VaultFeeHook.sol).
  • Test on 15,452 real 8h Binance funding observations (ETH, BTC, 2019 to 2026) as a proxy: maas has no funding history yet.
  • Pre-registered train/test split at 2023-07-01. Train-selected parameters, stationary block bootstrap (2000 resamples).
  • Rules compared: always-hedged, N-consecutive (the repo's design), EWMA with hysteresis.

Results​

  • Out of sample, no rule beats always-hedged at any cost tested (5 to 50 bps). At 35 bps (fee-paying comparator) the N-rule is significantly worse.
  • Negative funding is too transient: persistence p ≈ 0.5 against a required p* ≥ 0.94 to pay for a single switch. Break-even needs roughly nine days of negative funding.
  • Small N destroys yield: N=1 switches about 149 times a year, taking ETH from +10.9% to -0.7% p.a. Do not deploy N < 6.
  • Leverage on the short leg is the dominant risk. At 4x there were no liquidations in seven years. Liquidations appear from 6.7x and explode past about 10x.
  • Idle USDC yield barely reopens the question: at 4% p.a. the ETH EWMA rule gains +0.27 p.p. out of sample.

Design implications​

  1. Default to always-hedged. If a safety switch is wanted, use slow EWMA (h ≈ 48 periods, θ ≈ 1e-5): about 0.6 flips per year, never significantly harmful.
  2. Keep short leverage at or below 7x and keep spot gains in perp margin.
  3. Mechanism risks found in the audit: no time bound on the per-call funding clamp, no single-instance keeper lock (a duplicate keeper could charge several periods in one block), and liquidation reward makes latency a protocol variable.

Limits​

CEX funding is a proxy for maas's hourly, clamped funding. Basis mark-to-market, slippage and the vault's own funding impact are omitted. Liquidation counts are lower bounds. About 35 configurations per asset were tried, which favours the switching rules, and they still did not win.

:::note Impact on the demo The demo and idea use a short window (3h) and flip to FLAT on negative funding. That is a demo setting. The research above argues against it for production. :::