OpenMarket: A Synchronized Polymarket-Binance Dataset for High-Frequency Prediction-Market Research

Prediction markets have evolved as a mechanism to aggregate dispersed information through tradable contracts, with Polymarket exemplifying decentralized binary outcome markets. Prior microstructure research focused on centralized exchanges like NASDAQ, with datasets such as LOBSTER and FI-2010 providing high-frequency order book data. However, decentralized prediction markets, especially in crypto, present unique microstructure features—such as one-tick spreads, asymmetric information flow, and blockchain-specific delays—that remain underexplored. Existing studies have analyzed tick-level order book evidence or proprietary arbitrage data but lack publicly available, synchronized, millisecond-resolution datasets. This gap hampers comprehensive microstructure analysis, causal inference, and arbitrage modeling in decentralized crypto prediction markets. The current work addresses this by releasing a rigorously synchronized, high-frequency paired dataset, enabling detailed microstructure and timing studies, and fostering transparency and reproducibility in this emerging domain.

The core challenge lies in accurately synchronizing and pairing events across decentralized venues with inherently asynchronous clocks and network delays. Traditional microstructure datasets assume centralized, tightly synchronized exchanges, which do not apply to blockchain-based prediction markets. The lack of precise timing impairs causal inference, delay quantification, and arbitrage detection. Moreover, the high volatility and microstructure peculiarities—such as one-tick spreads and asymmetric information flow—compound the difficulty. Existing methods either rely on imperfect synchronization or coarse time windows, limiting the fidelity of microstructure analysis. Addressing these issues is crucial for understanding the true speed of information transfer, market efficiency, and the potential for high-frequency arbitrage in decentralized crypto markets.

The study introduces several key innovations: 1) a Rust-based multi-source WebSocket data collection pipeline capturing real-time market events with source and ingest timestamps; 2) a clock offset validation method quantifying maximum clock bias (~±99ms) and drift (≤6ms) over the dataset; 3) a 750ms windowed nearest-neighbor matching algorithm for event pairing, enabling precise lead-lag analysis without requiring perfect synchronization; 4) causal inference via event-based analysis, matching Binance large moves with Polymarket quote responses, independent of cross-venue timestamp alignment; 5) a walk-forward logistic regression model calibrated with Platt scaling, evaluated on a large, deduplicated dataset. These innovations collectively improve the reliability, granularity, and reproducibility of high-frequency microstructure analysis in decentralized prediction markets.

• �� Data collection: WebSocket collectors from Binance and Polymarket capture trades, order book updates, and events, storing both source and ingestion timestamps. • Synchronization validation: measuring maximum transport delays (~16ms for Binance, ~7ms for Polymarket), establishing bounds on clock drift (≤6ms) and bias (~±99ms). • Event pairing: within a 750ms window, nearest neighbor matching aligns Polymarket ticks with Binance trades, assigning quality flags based on lag magnitude. • Causal analysis: large Binance moves (≥5bps) are matched with Polymarket bid changes within 2s, measuring quote response lag (median 347ms) on collector timestamps. • Model training: features include order book metrics, returns, volatility; models are trained via walk-forward logistic regression with Platt calibration, evaluated with AUC and Brier scores. • Results: microstructure features show limited predictive power, confirming high market efficiency. • Infrastructure: Rust pipelines, Parquet exports, and validation scripts ensure reproducibility.

The experiments utilize data from Feb to May 2026, covering 54 Polymarket days and 57 Binance days. Features include spread, bid-ask imbalance, returns, volatility, and candle metrics. Models are trained with walk-forward validation, with hyperparameters tuned per market. Baseline comparisons involve naive Polymarket mid probabilities. Ablation studies test feature importance and pairing window sensitivity. The primary evaluation metrics are AUC, Brier score, and calibration error. Additional causal event analysis assesses market response delays. All experiments are conducted in high-performance computing environments to ensure reproducibility and robustness.

The models achieved an out-of-sample AUC of 0.8377, slightly below the naive market prior (0.8405). Microstructure features, such as one-tick spreads and lead-lag distributions (median 16ms), showed limited predictive advantage. Clock offset validation confirmed a maximum bias of ±99ms, supporting timing reliability. Causal event analysis demonstrated that large Binance moves influence Polymarket quotes after a median of 347ms, indicating a measurable but delayed information transfer. These findings reinforce the market’s high efficiency and highlight the challenges of short-term prediction using microstructure features.

The dataset enables detailed microstructure and timing analysis in decentralized prediction markets, supporting research on market efficiency, arbitrage opportunities, and high-frequency trading strategies. Industry practitioners can leverage this infrastructure for developing latency-aware algorithms and risk management tools. The methodology can be extended to other crypto venues and prediction markets, fostering transparency and reproducibility in high-frequency market analysis.

The synchronization relies on a single vantage point, with an estimated bias of ±99ms, limiting ultra-short-term inference accuracy. Models did not surpass market-implied probabilities, indicating microstructure features alone are insufficient for profitable prediction. High-frequency data collection is computationally intensive and sensitive to network delays, requiring further optimization. Future improvements include multi-source clock calibration, advanced models, and broader venue coverage.