Deterministic Off-Hours Reference Pricing & Source-Quality Guard for High-Beta US Equities (NVDA & TSLA)

MarketBridge solves the critical solvency and reference pricing vulnerability that occurs when primary US equity exchanges and ATS venues (CME, IEX, Blue Ocean ATS) are closed. During off-hours and weekend trading windows, traditional equity feeds go dark while on-chain perps and tokenized equities trade 24/7. MarketBridge ingests receipt-ordered multi-venue evidence, enforces family-level consensus, quarantines anomalies, and produces mathematically calibrated reference envelopes.

1. Family-Level Corroboration: Multiple resellers or repackagers of the same underlying quote are not independent evidence. MarketBridge categorizes sources into strict cryptographic families (hyperliquid_oracle, xstocks, ondo, perps_ats) and mandates $\ge 2$ independent, fresh families agreeing before admitting a price jump into the reference estimator. Single-venue spikes are quarantined immediately.
2. Strict Receipt-Order Processing: Events are evaluated strictly by arrival timestamp (received_at), guaranteeing that past states are never contaminated by future knowledge and eliminating race conditions.
3. Graceful Abstention over Stale Fallback: When corroboration is lacking or feeds drop out, the engine returns INSUFFICIENT_EVIDENCE with dynamic band widening, explicitly refusing to publish stale marks or hallucinatory predictions.
4. Venue Mark Isolation: Venue mark prices ($P_{mark}$) are captured strictly as un-admitted comparators to detect basis dislocations. Mark prices are mathematically barred from entering the independent reference estimator (oraclePx only).

MarketBridge replaces synthetic circularity with real external market evidence, empirical overnight factors, and distribution-free conformal bands:

• Live Weekend Adapters:
  • backend/marketbridge/adapters/hyperliquid.py: Runtime DEX deployer discovery via {"type":"perpDexs"}, indexing oraclePx for reference estimation and isolating markPx as comparator only. Assertions enforce exact index matching across universe and asset contexts.
  • backend/marketbridge/adapters/solana_tokens.py: High-throughput Jupiter Lite Price API (v3) with runtime Base58 mint validation and separate family segregation for xStocks and Ondo.
  • Health dropout transitions emit explicit DROPOUT states to prevent silent failures.
• Real Historical Incident Defense (fixtures/skhynix_20260728.jsonl):
  • Reconstructed the documented SK Hynix pre-market anomaly of 28 July 2026 (08:00 KST), where an uncorroborated single-venue error printed 1,272,000 KRW (~$868, -29.96% below the 1,816,000 KRW prior close) before reverting within 2 minutes.
  • MarketBridge quarantined the bad print on receipt, preserving simulated paper solvency while an unguarded benchmark suffered immediate false liquidation.
• Adversarial Robustness Sweep (scripts/adversarial_sweep.py):
  • Swept shock magnitudes from 0.5% to 30% across single vs. multi-family distributions and 0–12s latency.
  • Results logged to artifacts/robustness.json with an empirical ROC curve generated in artifacts/roc_curve.png. Confirmed genuine detection boundaries at subtle sub-threshold shocks (<3%).
• Overnight-Specific Factor Beta (scripts/fit_beta.py):
  • Evaluated 504 trading days of close-to-open gaps (distinguishing IEX prints from primary opening auctions).
  • Fitted overnight Ridge regression shrunk toward sector mean ($\beta_{sector} = 1.15$), adhering to Hendershott, Livdan & Rösch (JFE 2020) principles on overnight vs. intraday factor sign separation.
  • Persisted to artifacts/fitted_betas.json with cryptographic hash verification ($\beta_{NVDA} = 1.1502, \beta_{TSLA} = 1.1506$).
• Split-Conformal Uncertainty Bands:
  • Dynamic envelope replaces static heuristic spreads with split-conformal residual quantiles scaled by EWMA realized volatility and staleness penalties: $$\text{Band Half-Width} = q_{conformal} \times \sigma_{EWMA} \times (1 + 0.05 \cdot \text{age})$$
  • Ladder evaluation (scripts/evaluate.py) tests Last-Close, Factor-Baseline, and MarketBridge across 80%, 90%, and 95% target coverage, outputting artifacts/evaluation_ladder.json and artifacts/reliability.png.

   ┌─────────────────────────────────────────────────────────────┐
   │             Live 24/7 Weekend Ingestion Layer               │
   │  ┌──────────────────────┐  ┌─────────────────────────────┐  │
   │  │  Hyperliquid L1 Info │  │   Solana Jupiter Lite API   │  │
   │  │  oraclePx (Evidence) │  │  xStocks (NVDAx / TSLAx)    │  │
   │  │  markPx (Comparator)│  │  Ondo US Yield / Equities   │  │
   │  └──────────┬───────────┘  └──────────────┬──────────────┘  │
   └─────────────┼─────────────────────────────┼─────────────────┘
                 │                             │
                 ▼                             ▼
   ┌─────────────────────────────────────────────────────────────┐
   │                  MarketBridge Core Engine                   │
   │   • Strict Arrival-Time Receipt Ordering                    │
   │   • ≥2 Independent Family Consensus Guard                   │
   │   • Overnight Ridge Factor Adjustment (β_overnight)         │
   │   • Split-Conformal Volatility-Scaled Uncertainty Band      │
   │   • Unverified Jump Isolation & Health Dropout Handling     │
   └─────────────────────────────┬───────────────────────────────┘
                                 │
                 ┌───────────────┴───────────────┐
                 ▼                               ▼
   ┌───────────────────────────┐   ┌───────────────────────────┐
   │  FastAPI Analytics Engine │   │ Next.js Terminal (CMC UI) │
   │  • /health                │   │ • Lightweight Charts v5   │
   │  • /v1/demo/scenarios     │   │ • Source Family Quorum    │
   │  • /v1/demo/evaluation    │   │ • HL Mark-vs-Oracle Panel │
   │  • /v1/live/sources       │   │ • Conformal Widening Pill │
   └───────────────────────────┘   └───────────────────────────┘

• Python 3.12+
• Node.js 20+ & npm
• uv and make

# Sync dependencies
uv sync --frozen
npm --prefix apps/web ci

# Run verification suite (pytest, ruff lint, tsc typecheck, next build)
make verify

# Start unified demo (FastAPI on 8000 serving pre-built Next.js frontend)
make demo

Open http://localhost:8000 in your browser.

For live development with hot module replacement:

# Terminal 1: Backend API
make serve

# Terminal 2: Next.js Frontend
make dev

Open http://localhost:3000.

# Run three-way baseline ladder evaluation (Last-Close vs. Factor vs. MarketBridge)
python scripts/evaluate.py

# Run adversarial shock sweep (0.5% to 30%) and generate ROC curve
python scripts/adversarial_sweep.py

# Fit overnight-specific beta on close-to-open gaps
python scripts/fit_beta.py

# Capture a frozen snapshot of all live weekend sources for 1-keystroke demo fallback
python scripts/capture_fixture.py

# Replay specific scenario and export trace
SCENARIO=bad-print SYMBOL=NVDA make replay

Scope & Operating Parameters: MarketBridge is an off-hours reference pricing and source-quality risk guard designed for non-clearing trading windows when primary exchange books are dark. Live weekend feeds ingest decentralized L1 perp oracles (Hyperliquid) and Solana tokenized equities (xStocks, Ondo); during primary exchange hours, exchange auctions and ATS feeds take precedence. MarketBridge does not execute customer trades or take custody of assets; it produces deterministic reference intervals, anomaly quarantines, and solvency state vectors for risk systems and automated liquidators.

MarketBridge is open-source software licensed under the MIT License.