RainBot review
Our verdict
Technically the closest thing to WeatherBot in the market, with a near-identical model stack and an unusually good telemetry dashboard. Its 30-day numbers are specific and come with a fill count, which we rate highly, and its whitepaper is one of the few in this category that shows the arithmetic rather than gesturing at it. It loses ground on the missing fee disclosure — asking users to pair a wallet before they can find out what they will be charged is the wrong order — and on an anonymous operator. Fixing the first would move this up the table.
How it works
RainBot streams Polymarket's weather order book over WebSocket and fuses five institutional forecast models — GFS, ECMWF, UKMO, NWS and HKO, the last of which is the official gauge that settles Hong Kong markets. A model whose forecast deviates more than 1.5 standard deviations from the raw mean has its weight halved, so one bad model cannot drag the ensemble. A 10-year NOAA NCEI prior is folded in by Bayesian update, and where forecast and climatology disagree by more than 1.5σ the combined σ is widened by 30%. The edge threshold floats between 3% and 8% with forecast confidence; sizing is quarter-Kelly clamped to 15% of bankroll. Claude and Gemini each review a candidate with the full briefing — model spread, climatology, order books, history — and return BUY_YES, BUY_NO or SKIP before anything is submitted. Execution runs under two seconds from tick to order, using signature type 2 (POLY_GNOSIS_SAFE) on Polygon, fill-or-kill first with a GTC limit at 2% slippage as fallback. Exits are five-way: 60% of edge captured, edge collapse below 2%, a 40% trailing stop, 15% drawdown, and time decay inside two hours of resolution.
Strengths and weaknesses
What works
- Publishes a precise 30-day win rate (71.4%) with the fill count behind it — 7,184 fills is a sample size you can reason about.
- Per-member P&L and the average are both disclosed, so the aggregate figure is not hiding a single outlier account.
- Documented non-custodial model: the private key is used locally to derive the address and sign the Polymarket credential trio, and is stated never to cross the network or be persisted server-side.
- The whitepaper gives the actual arithmetic — ensemble weighting, the Bayesian prior, the Kelly clamp — rather than describing the strategy in the abstract.
- Outlier suppression when forecast models disagree is a sound guard, and the failure paths (dead order books skipped, FOK then GTC fallback, execution lock per token) are written down.
What doesn't
- No pricing disclosed anywhere on the public site. This is the single biggest gap and the main reason it sits behind WeatherBot in our index.
- 986 active traders is a modest base; per-member figures are more sensitive to a few strong accounts than they would be at scale.
- Weather-only scope, and an undisclosed dependency on upstream forecast APIs — nothing states what happens to open positions if a model feed goes dark.
- No team, company or legal entity is named anywhere on the site.
- Nothing is published on outage recovery or state restoration after an interruption, which is the operational question that matters most for a bot holding live positions.