Oracle3
Autonomous prediction market trading agent across Kalshi, Polymarket, and Solana.
Why this exists
Prediction markets price binary contracts at systematically biased levels — a true 50/50 contract typically trades around 0.57 (favorite-longshot bias, $\hat{\lambda} \approx 0.183$). Most trading bots ignore this distortion entirely. Oracle3 operationalizes a peer-reviewed pricing model, calibrated on 291,309 resolved contracts across six venues, to systematically harvest the bias through arbitrage detection and Kelly-sized model trades.
This system deploys the exact $\lambda$ estimates and covariate model from prediction-market-pricing (Yang, 2026) as its real-time pricing engine.
How oracle3 differs from existing prediction-market tools
| Oracle3 | polymarket-whales | prediction-market-maker | py-clob-client | |
|---|---|---|---|---|
| Pricing model | Wang Transform (calibrated MLE) | None | Bid-ask MM | None |
| Constraint-based arbitrage | 8 strategies | None | None | N/A |
| Multi-venue | Kalshi + Polymarket + Solana | Polymarket only | Polymarket only | Polymarket only |
| On-chain execution | Solana via DFlow + Jito | No | No | N/A (SDK) |
| Working paper | Yang (2026), SSRN | No | No | No |
| Tests | 633 | 0 | 0 | 50+ |
| License | Apache 2.0 | MIT | MIT | MIT |
Architecture
graph TD
A[Wang Transform Pricing Engine<br/>MLE coefficients from paper] --> B[Fair Value Estimator<br/>Model Greeks · Kelly Sizing]
B --> C[Strategy Layer]
C --> D[8 Constraint-Based Arbitrage]
C --> E[2 Model-Driven Strategies]
C --> F[LLM Agent Strategies]
D --> G[Trading Engine<br/>SpreadExecutor · Risk Manager · Position Tracker]
E --> G
F --> G
G --> H[Kalshi]
G --> I[Polymarket]
G --> J[Solana / DFlow]
Strategies
Constraint-based arbitrage — each exploits a violated probability axiom:
| Strategy | Invariant |
|---|---|
| Cross-Market | Same event, same price across exchanges |
| Exclusivity | $P(A) + P(B) \leq 1$ for mutually exclusive events |
| Implication | $P(A) \leq P(B)$ when A implies B |
| Conditional | $P(A \mid B) \in [L, U]$ within derived bounds |
| Event Sum | $\sum P(\text{outcome}_i) = 1$ within an event |
| Structural | $P(A) = \beta \cdot P(B) + \alpha$ from calibrated model |
Statistical arbitrage: cointegration spread (self-calibrating z-score), lead-lag (cross-correlation).
Model-driven: fair value divergence (Wang-model edge), premium decay (rides predictable premium lifecycle).
Pricing Engine
Deploys the Wang Transform from Yang (2026), calibrated on 291,309 contracts across 6 platforms:
$$p^{\text{mkt}} = \Phi\bigl(\Phi^{-1}(p^*) + \lambda\bigr), \quad \hat{\lambda} = 0.183 ; (p < 10^{-15})$$
- Hierarchical model: $\lambda_i = 0.259 - 0.072 \ln(1+V) + 0.143 \ln(1+D) - 0.477 |p-0.5|$
- Model Greeks: $\partial p / \partial \lambda$, Kelly fraction, edge decay rate
- Online calibrator: hybrid batch MLE + streaming EWMA with category shrinkage
- Correlation-aware risk: EWMA correlation matrix, effective exposure limits
Yang, Y. (2026). Pricing Prediction Markets: Risk Premiums, Incomplete Markets, and a Decomposition Framework. Working Paper, UIUC. [Replication package]
Quick Start
git clone https://github.com/YichengYang-Ethan/oracle3.git && cd oracle3
poetry install
oracle3 market list --exchange polymarket --limit 10
oracle3 dashboard --exchange solana --initial-capital 10000
See docs for full CLI reference.
Key Technical Choices
- Event-driven async engine with snapshot persistence and Unix socket control (pause/resume/killswitch)
- SpreadExecutor with automatic LIFO unwind on partial fills — no naked multi-leg positions
- Dual-layer risk: local position/drawdown/exposure limits + Solana
simulateTransactionpre-flight - On-chain audit trail via Solana Memo program; Jito bundle submission for MEV protection
- 633 tests, ruff, mypy, codespell CI on every push
Star History
Contributors
If oracle3 helps your research or trading, please ⭐ star the repo — it helps others find it.
License
Apache 2.0 — see LICENSE for details.
This software is for research and educational purposes. Trading involves financial risk.