Skip to content

2026 - 06

Goal: Confirm or kill MR. Find a second edge on a surface where MM actually works. | Phase: 1 (Validate)

May closed two of three prop-spot families (passive MM, price-trend) with first-principles verdicts and left MR as the one live thread. June is a pivot: stop iterating dead families, prove or disprove MR, and chase the second edge where it's cheapest to find — MM on a wide-spread surface (where the adverse-selection math flips positive), reusing the MM machinery we already built. Carry/funding is a promising but infra-heavy axis; June only explores and quick-validates it — no build commitment until the cheap data work says it's worth it.

1. Strategy: Confirm or Kill MeanReversion

MR is the survivor — live-positive (+26 to +41 bps, 100% win) but on ~3 independent events, and unfalsifiable in the kline backtest (blind to its 2s intra-bar entries). Get it to a sample that can decide.

  • Extend MR live to a decisive sample — more independent events, genuinely independent setups (not baseline + tight_sr firing on the same trigger)
  • aggTrade-faithful sim path for MR (the only offline route that can see its entries) — or accept live as sole ground truth and size accordingly
  • Gate MR with the VSR regime classifier — enter only in RANGING/contraction regimes; stand down in predicted expansion. First live use of an ML output in a strategy.
  • Decision by month-end: MR is candidate #1 (convergent + margin-positive) or it joins the dead pile

2. MM: Find the Surface Where It Works (the second-edge bet)

The MM machinery (v9-v15) is built and proven to run; it just has no edge on OKX BTC/ETH (spread 0.01-0.05 bps, adverse selection beats the −0.75 rebate). MM economics flip positive where the spread is wide enough. Two fast, machinery-reusing surfaces to test — this is June's most likely route to a second profitable strategy.

  • Binance.US canary (fast + cheap): small gateway mod (BINANCE_US, ≈ Binance-global API shape), one cross-venue-informed MarketMakingTrueValue instance. Wide spread (0.5-2.6 bps) + real global→US lead. Volume-capped (~$100/day ceiling) so it's a capability + latency testbed, not P&L — but it's the cheapest place to learn the reprice-latency → adverse-selection curve, and positions us pre-launch on the shared-liquidity catalyst.
  • Wider-spread alt symbols (the actual P&L shot): pre-gate by confirming spread > ~2 bps on a candidate (HYPE / SOL / similar) from an observe-only stream before wiring trading. If it clears, wire order rules + one MM cohort. No derivatives, no new infra class — reuses everything.
  • Deliverable either way: the per-style reprice-latency → adverse-selection curve, and a clear read on whether any accessible spot surface gives MM positive per-fill economics.

3. ML: VSR Integration + Direction-Tuned Features

VSR is production-grade (MCC 0.40) but unused. Put it to work; push the one remaining directional lever.

  • Wire the VSR gate into the strategy layer as routing/sizing (first consumer: MR, §1)
  • E6/E7 direction-tuned features (return lags, RSI, MACD, funding) per regime — the non-price/multivariate axis. Strict OOS; treat any in-sample edge as guilty.
  • AdaptiveDirectional v3 — Vicky's first strategy-repo contribution (onboarding #55); design discussion before code.

4. Carry: Explore + Quick-Validate Only (no build commitment)

Carry/funding is a genuinely new edge class with a better epistemic position (directly observable) — but it's an infra-heavy lift (perp support, funding feed, cross-instrument margin) gated on jurisdiction, and its viability for us is unproven. June does the cheap data work to decide whether it's even worth pursuing. No engine perp build this month.

  • Funding-dispersion screen (data-only, days, zero engine work) — live funding across accessible venues + the alt universe, ranked by net-of-cost cross-venue spread. Sizes the prize. Majors are competed (~2% net); the question is whether accessible alt dispersion clears costs + tail risk.
  • Historical carry backtest, per-regime — realized carry net of honest cost + tail haircut, per quarter. Stress-survival test (same validation rigor as the directional memo).
  • Output: a go/no-go recommendation on the carry axis — not a build. If it looks real, then scope the engine perp work as a July decision. If not, shelve it with a memo like the others.
  • Separate low-risk parallel track (not the HFT carry book): holdings yield — stake ETH (~3-4%) / lend BTC on the long-term treasury. Real yield on idle capital, strictly walled off from any delta-neutral carry strategy (firm principle: never short perp against the treasury).

Strategic note: carry is the class Windhorse/Cody run. If validation is positive, a future carry v1 doubles as a credibility artifact for the RL VC pipeline — a reason to weight it up if the economics are there. Not a reason to pre-commit before they're proven.

Team Focus

MJ: MR confirm/kill, MM surface hunt (Binance.US + alt wiring), carry validation (screen + backtest), VSR gating.

Vicky: VSR integration, direction-tuned features, AdaptiveDirectional v3.

Gaddafi: Frontend — latency/markout instrumentation for the MM testbed; continued dashboard work.

Phase Status

Phase 1 (Validate) — re-scoped. No longer "iterate the families to convergence" (two dead). The question is: confirm MR as candidate #1, and find a second edge — most likely MM on a wide-spread surface.

June success: (1) MR confirmed-or-killed with a decisive sample; (2) a clear read on whether any accessible spot surface gives MM positive economics; (3) a go/no-go on carry from cheap validation. Exiting Phase 1 still needs 2 strategies with edge — June is about locking the first (MR) and identifying the second, not yet proving it at scale.