| title | Oracle3: An open-source autonomous trading agent for prediction markets with a Wang Transform pricing engine | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| tags |
|
|||||||||
| authors |
|
|||||||||
| affiliations |
|
|||||||||
| date | 7 May 2026 | |||||||||
| bibliography | paper.bib |
Prediction markets price binary contracts at systematically biased levels: a contract whose objective probability is 50% trades on average around 57 cents, an empirical regularity known as the favorite-longshot bias [@thaler:1988]. Despite a long literature documenting this distortion, most open-source trading bots ignore it, treating market prices as unbiased estimates of probability. Oracle3 is a Python framework that operationalizes a peer-reviewed risk-neutral pricing model for binary outcome contracts and uses it to drive automated trading across multiple venues (Kalshi, Polymarket, and Solana-based DFlow). At its core is a Wang Transform [@wang:2000; @wang:2002] calibrated by maximum-likelihood estimation on 291,309 resolved contracts spanning six platforms, with hierarchical covariates for volume, days-to-expiry, and contract moneyness. The library exposes the model as a fair-value engine and pairs it with eight constraint-based arbitrage strategies, statistical-arbitrage strategies, model-Greek-driven sizing, and a risk manager — all wired into an event-driven async trading core with snapshot persistence, killswitch support, and on-chain audit trails.
Quantitative research on prediction-market pricing has accelerated, but the gap between empirical pricing literature and reproducible, deployable software remains wide. Existing open-source bots (e.g., polymarket-py, kalshi-python) are thin client wrappers; they do not estimate fair value, do not operationalize known pricing distortions, and offer no shared abstraction across venues. Conversely, academic replication packages typically stop at calibration and rarely produce a runnable trading agent. Oracle3 closes this gap by shipping the maximum-likelihood estimates from @yang:2026 directly as a real-time pricing engine, exposing the same hierarchical covariate model used in the paper, and connecting it to live order-book data, execution, and risk management.
The intended audience is twofold. For researchers studying prediction-market efficiency, the favorite-longshot bias [@thaler:1988; @snowberg:2010], and risk-neutral pricing [@harrison:1979; @wang:2000], Oracle3 provides a reproducible, well-tested implementation of a multi-venue pricing-and-trading pipeline that can be used as a baseline for new strategies or as a calibration testbed. For practitioners, it provides a transparent, auditable agent whose every signal is grounded in an explicit pricing model rather than opaque heuristics.
Several Python tools touch the prediction-market space, but each is partial. py-clob-client and kalshi-python are vendor SDKs without modeling. The crypto-trading-bot family targets centralized exchanges and currency markets, not binary outcome markets. Research code accompanying empirical-finance papers is typically Jupyter-driven, focused on calibration rather than deployment. Outside Python, decentralized-finance projects on Solana such as dflow-program provide on-chain primitives but no off-chain modeling layer. Oracle3 is, to our knowledge, the first open-source library that combines (i) calibrated probabilistic pricing for binary contracts, (ii) constraint-based arbitrage detection across heterogeneous venues, and (iii) production-grade execution with on-chain settlement.
Oracle3 is organized as four layers:
-
Pricing engine (
oracle3/pricing/). Implements the Wang Transform$p^{\mathrm{mkt}} = \Phi(\Phi^{-1}(p^*) + \lambda)$ with$\hat\lambda = 0.183$ at the global level and a hierarchical covariate model$\lambda_i = 0.259 - 0.072 \ln(1+V) + 0.143 \ln(1+D) - 0.477 |p - 0.5|$ where$V$ is daily volume,$D$ is days to expiry, and$|p-0.5|$ is moneyness. An online recalibrator combines batch MLE with a streaming EWMA estimator and category shrinkage, so the model adapts to regime change without forgetting prior data. -
Strategy layer (
oracle3/strategy/). Eight constraint-based strategies each enforce a probability-axiom invariant (cross-market identity, exclusivity bound$P(A)+P(B)\le 1$ , implication monotonicity, conditional bounds, event-sum unity, structural relationships from the calibrated model) and emit signals when violations exceed thresholds. Two model-driven strategies trade fair-value divergence and a "premium decay" lifecycle effect. Statistical-arbitrage strategies (cointegration spread, lead-lag) provide complementary signals. -
Trading core (
oracle3/core/,oracle3/trader/). An async event loop dispatches strategy signals to aSpreadExecutorthat posts multi-leg orders atomically and unwinds via LIFO on partial fills, eliminating naked legs. A dual-layer risk manager enforces position, drawdown, and exposure limits locally, and additionally calls SolanasimulateTransactionfor pre-flight on-chain checks. -
Infrastructure (
oracle3/cli/,oracle3/data/). A Click-based CLI exposes market discovery, live dashboards, backtests, and paper-run modes. Snapshot persistence and a Unix-socket control plane enable pause/resume/killswitch operations on long-running agents. Solana submissions go through Jito bundles, with a Memo-program audit trail for every fill.
The codebase is type-checked with mypy, linted with ruff, and validated by 633 unit tests and integration tests on every push, enforced by GitHub Actions. The design has been used to run paper-trading sessions on Polymarket and Kalshi, and the pricing engine is shared with the replication package of @yang:2026.
The author thanks the maintainers of the Polymarket, Kalshi, and DFlow APIs for documentation; the prediction-market research community whose empirical work motivates this software; and contributors who reported issues and suggested strategies. Any errors are the author's own.