FedWatch Replication vs. Polymarket

Built a from-scratch replication of CME's FedWatch methodology — converting Fed funds futures (ZQ contract) prices into market-implied probabilities for FOMC rate decisions to answer a question CME's own tool doesn't: how does that probability compare against an independent market pricing the same event? Polymarket runs prediction markets on FOMC decisions, so this project builds the FedWatch-style engine from raw futures data, validates it against CME's published methodology, and compares its output to Polymarket's prices over time.

At the core is a deconvolution engine that decomposes each contract's implied rate into an integer + mantissa step per CME's published method, then convolves that forward across the full meeting sequence into a probability distribution per meeting which I then regression-tested exactly against CME's own worked example. Polymarket's FOMC-related markets are matched to meetings through a manually reviewed candidate table, since automated question-text matching alone isn't reliable enough to trust blindly.

Highlights

• Deconvolution engine reproducing CME's own worked example exactly, independent of any external data file.
• 66-test suite that runs with no network access, backed by CSV fixtures.
• Measured across 21 settled FOMC meetings (2023–2026): a lower Brier score than Polymarket in 15/21 meetings, and a statistically robust ~1–3 day lead over Polymarket's pricing, confirmed by two independent methods (threshold-crossing and cross-correlation, p < 0.01).
• A backtest of the resulting trading rule with zero losing meetings, treated as a red flag rather than a result, and traced to a specific mechanism (the position that survives to resolution wins 21/21) rather than taken as evidence of skill.
• Optional live notifier that checks daily for a qualifying signal and pushes a Telegram alert, run over local cron rather than GitHub Actions after investing.com started returning 403s specifically to GitHub's runner IP ranges.