Polymarket bot
Machine learning and data flow for prediction-market trading

142 / Polymarket mechanics, order-book microstructure, and a practical workflow for building trading bots with reproducible blockchain data. Covers DuckDB plus Parquet pipelines, feature engineering, backtesting discipline, trigger-based trade ops, latency tradeoffs, and Kelly sizing ideas for prediction-market strategies.

Timeline

00:00 Intro

00:49 Why 2026 is the time to start

02:12 Polymarket order process

06:26 CLOB market mechanism

12:12 Data storage with Parquet

15:48 Raw and derived data

23:38 Data features and pipeline

26:49 Thesis backtesting

32:53 Trade ops

37:27 Engineering with AI

45:13 Connect digital and physical worlds

Participants

@fulldecent

William Entriken

@

@

@

Episode notes

• Polymarket
• Polymarket docs
• DuckDB
• Apache Parquet
• Polygon
• Kalshi
• Augur
• Kelly criterion explainer

Prediction-market trading is used here as a concrete way to learn modern data and ML workflow discipline, not as a promise of easy money.

Core walkthrough

• The talk starts with Polymarket mechanics: yes/no outcome tokens, wallet signing, bundling, and final settlement on Polygon.
• A major theme is data provenance. The recommended source of truth is chain data instead of platform-provided historical exports.
• Market microstructure is explained with maker/taker behavior and CLOB matching, including why trade direction and sequence matter for feature design.

Data architecture and tooling

• The suggested stack is DuckDB with Parquet files for full-table analytical transforms.
• The speaker contrasts row-oriented OLTP thinking with columnar OLAP workflows for large historical backtests.
• A repeated implementation pattern is immutable partitions plus manifests and hashes for reproducibility and forensics.

Strategy and backtesting themes

• Multiple thesis families are discussed: range and ping-pong market making, account-following, anomaly detection, buy-and-hold timing, and cross-market arbitrage.
• External signals can be useful, but only when they are timestamped to avoid look-ahead leakage.
• Feature ideas include account behavior profiles, trade timing distributions, exposure measures, and volatility bands.

Trade operations and execution

• Production logic is framed as trigger-driven state transitions with detailed logging of conditions and execution context.
• The conversation highlights slippage and latency constraints, especially for short-horizon markets where seconds can matter.
• Position sizing is discussed through the Kelly criterion as a guardrail against overbetting.

Engineering process with LLMs

• The workflow recommends drafting hypotheses in plain language first, then defining API contracts before code generation.
• Testing is split into both code-level unit tests and data-shape tests.
• Frequent commits and reproducible runs are treated as mandatory for debugging model drift and execution failures.

Open items for human follow-up

• Confirm youtube-full URL.
• Confirm discussion URL.
• Confirm whether additional participants should be credited.