PolymarketBayesianCandlestickKelly

This strategy (tylervsholden-source/polymarket-bot) combines three distinct quantitative methods: Bayesian inference for updating event probabilities as new information arrives, classic candlestick pattern recognition applied to Polymarket price series, and Kelly criterion for position sizing. The Bayesian component treats each market's implied probability as a prior and updates it with observed price action, volume, and external signals. Candlestick patterns (doji, engulfing, hammer, etc.) are detected on the price chart of the YES/NO shares to identify short-term reversals or continuations. Kelly sizing then determines optimal bet size based on the estimated edge and variance.

Entry signals are generated when the Bayesian posterior probability diverges from the market price by a threshold AND a bullish/bearish candlestick pattern confirms direction. Exit occurs when the posterior converges to market price, when a stop-loss is hit, or at a fixed time horizon. The bot is designed to trade all Polymarket categories, not just crypto, giving it broad applicability.

Implementation: The repo is 6.9MB of Python code, created in March 2026 and updated September 15, 2026. It includes modules for data ingestion, pattern detection, Bayesian updating, and order execution (though execution may be paper-traded by default). The use of Kelly sizing suggests a focus on long-term bankroll growth rather than high-frequency scalping.

Risks: Bayesian models require careful prior selection and can be slow to adapt to regime changes. Candlestick patterns are less reliable on prediction markets than on equities due to lower liquidity and discontinuous price jumps. Kelly sizing can be overly aggressive if edge estimates are wrong, leading to large drawdowns. The repository has 0 stars and 0 forks, so community validation is absent.

Verified results: No public PnL, win rate, or backtest results are provided in the repository metadata. The strategy is theoretically sound but lacks empirical validation from the intelligence data.