Engine · AsPredicted #288615 · 2026-05-01

CME V0.2-Perps

Seven-constraint coherence-violation engine for crypto derivatives. Closes the “anchor process” gap named in MNX's 2026-03-25 “Why Perps?” post: a perpetual contract needs a credible index or settlement process to anchor against. Coherence violations are the quantitative signal that the anchor is broken.

AsPredicted #288615Filed 2026-05-01Freeze commit adb99d6V0.1 ancestor

The seven constraints

Each constraint enforces a no-arbitrage relation across the crypto-derivative graph. A persistent violation, after subtracting depth-aware execution costs, is a static or quasi-static arbitrage signal.

Cash-and-carry parity

perp − spot = funding basis over the funding interval

Canonical perp-spot anchor. Violation = arbitrageable basis between the perp and spot legs after expected funding cost. Default emission threshold: 50 bps cost-adjusted.

Triangle parity

Same-underlying perp pair vs stablecoin FX must be self-consistent

ETH-USDC perp / ETH-USDT perp must price-imply the same USDC/USDT FX, after fees. Violation = 3-leg synthetic-FX arbitrage. Default threshold: 30 bps cost-adjusted.

Put-call parity

C − P = S − K · exp(−r·T) on European options

Synthetic forward (long call + short put at the same strike) must equal the actual forward (spot minus discounted strike). Violation = 3-leg arb across call, put, underlying. Default threshold: 30 bps cost-adjusted.

Butterfly non-negativity (Carr-Madan generalized)

(K3 − K2)·C(K1) − (K3 − K1)·C(K2) + (K2 − K1)·C(K3) ≥ 0

Generalized for any K1 < K2 < K3 (no symmetry required). Reduces to the classical symmetric butterfly when K2 is the midpoint. Violation = static arb at expiry; non-negative payoff replicated for negative cost. Default threshold: 20 bps cost-adjusted.

Calendar-spread monotonicity (Litterman-Scheinkman PCA)

Forward funding curve decomposes into level + slope + curvature factors

Three-factor PCA on historical funding curves. Anomalous curvature factor (rolling z-score above threshold) signals a non-monotone “hump” in the term structure that is statistically unusual relative to the historical regime.

Cross-venue 4-corner coherence

DEX perp + CEX perp + options synthetic forward + spot reference must agree

Highest-EV constraint in the set: leverages cross-venue capital inefficiency that single-venue traders cannot exploit. Spread between max and min corner price (after depth-aware cost) is the violation magnitude. Default threshold: 60 bps.

Vertical-spread monotonicity

0 ≤ C(K1) − C(K2) ≤ K2 − K1 for K1 < K2

Both bounds enforce no-arb on adjacent-strike call options. Lower-bound violation: K2 call worth more than K1 call (impossible under no-arb). Upper-bound violation: call spread exceeds the strike differential. Per Carr-Madan 2005 Theorem 2 jointly with the butterfly constraint. Default threshold: 15 bps cost-adjusted.

Cost adjustment

Every constraint subtracts a depth-aware execution cost before emitting a signal. When the caller provides top-of-book spread + depth + target notional, the cost model uses a linear-impact slippage path (Almgren-Chriss 2001 + Cont-Kukanov-Stoikov 2014). When depth fields are unavailable, the model falls back to fixed fees + slippage per leg.

per_leg_entry_bps = fees_bps
                  + (top_spread_bps / 2)                       # half-spread crossing
                  + (top_spread_bps · depth_consumed_fraction) # linear-impact slippage
round_trip_cost = 2 · per_leg_entry_bps                          # entry + exit

Audit chain

Every emitted signal carries a deterministic UUIDv5 derived from a canonical-JSON SHA-256 hash of its violation provenance. Re-running the orchestrator on the same inputs produces identical signal IDs, making the emit ledger idempotent under replay. Resolution outcomes hash forward: prev_hash of each row equals the previous row's row_hash, forming a tamper-evident chain that anyone can verify in browser at /research/verify.

Pre-registered backtest

AsPredicted #288615 locks the methodology + hypotheses ex-ante. The 90-day walk-forward backtest runs after the daily emit cron has accumulated signal data and resolutions. Three formal hypotheses:

  • H1 (signal alpha): Net annualized Sharpe ratio > 1.0 after Hyperliquid + Binance + Deribit fees, depth-aware slippage, funding cost, and capital opportunity cost. Permutation null at p < 0.05.
  • H2 (capacity ceiling): Sum of modeled net EV < 50,000 USD/day on average. Bounds the trader-tier monetization ceiling.
  • H3 (per-constraint significance): Each of the seven constraint kinds contributes positive Sharpe individually with a 95% bootstrap CI excluding zero. Failing constraints get dropped from the production orchestrator.

References

  • Carr, P. and Madan, D. (2005). “A Note on Sufficient Conditions for No Arbitrage.” Finance Research Letters 2(3), 125-130. Equation (2) is the generalized butterfly form; Theorem 2 jointly characterizes butterfly + vertical-spread no-arb conditions.
  • Litterman, R. and Scheinkman, J. (1991). “Common Factors Affecting Bond Returns.” Journal of Fixed Income 1(1), 54-61. Level / slope / curvature factor decomposition.
  • Almgren, R. and Chriss, N. (2001). “Optimal execution of portfolio transactions.” Journal of Risk 3, 5-39. Linear-impact framework underlying the depth-aware cost model.
  • Cont, R., Kukanov, A., and Stoikov, S. (2014). “The price impact of order book events.” Journal of Financial Econometrics 12(1), 47-88.
  • Hull, J. (2021). Options, Futures, and Other Derivatives, 11th ed. Sections 5.6, 11.3, 11.5 (term structure, put-call parity, vertical-spread bounds).
  • López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley. Ch. 7 (purged k-fold) for fold geometry in the H1 backtest.
  • Convexly Coherent Markets Engine V0.1 (prediction-market ancestor) at /research/coherent-markets-engine-v1.
  • Convexly Edge Score V2-Perps composite (per-trader skill ranking; sister methodology) at /research/v2-perps.

This work used AI tools (Claude, GPT-4) as research aids during methodology design and pre-publication review. All claims on this page are reproducible from the freeze-commit code modules under services/api/app/engine/cme_v0_2_perps/ with the documented random seed. No claim on this page is taken as true on the basis of an AI tool's output; every quantitative result is recomputable from the freeze commit.