A new research paper posted on arXiv (q-fin.TR) documents a rigorous but unsuccessful attempt to generate a trading edge by linking Polymarket's prediction markets (BTC 15-minute binary contracts) to Binance BTC/USDT order flow. The study, titled OpenMarket: A Synchronized Polymarket-Binance Dataset for High-Frequency Prediction-Market Research, used 43 microstructure features — including order book imbalance, depth, volatility, and latency — in a walk-forward logistic model. Result: the model does not beat, and slightly underperforms, the probability already implied by Polymarket's own order book. In simulation, net trading generated -0.116 normalized payoff units per attempted trade after fees and slippage.
What this means for algorithmic traders and microstructure researchers
The failure of this model is no surprise to those working on trading signal calibration. It confirms a fundamental principle: when markets are liquid and order books are deep, price information is already embedded in spreads and imbalances. The fact that 43 features fail to extract net alpha after costs is an empirical validation of the relative efficiency of high-frequency prediction markets.
Strategy Arena: Smart Money Evolved Validation
At Strategy Arena, we observed a similar phenomenon when validating our Smart Money Evolved cross-asset metric. After Monte Carlo cross-validation filtering across 15 assets, this metric showed an ability to discriminate configurations where a microstructure signal can be exploited — and where it cannot. The OpenMarket study illustrates exactly this point: without rigorous out-of-sample validation and transaction cost accounting, a seemingly sophisticated model may be nothing more than overfitted noise.
Why this result matters for the anti-2CV community
The anti-2CV (cost, complexity, volatility, validation) approach rejects models that fail the test of practical robustness. Here, the model fails on all four dimensions: - Cost: fees and slippage erase any gross advantage. - Complexity: 43 features add nothing over the simple order book. - Volatility: 15-minute prediction markets are too volatile for a linear model to adapt. - Validation: out-of-sample walk-forward shows underperformance.
Caveat
This analysis is based on a backtest and paper trading simulation. Results are not proof of profitability in live conditions. Fees, latency, and liquidity can vary significantly. For a full understanding of our validation methodology, see our methodology page.
References - Original paper: OpenMarket: A Synchronized Polymarket-Binance Dataset for High-Frequency Prediction-Market Research - Strategy Arena metric: Smart Money Evolved cross-asset