LOWCELL

Methodology

What LOWCELL shows, where the data comes from, and what is not yet validated.

Data sources

  • Gamma API: event discovery, market metadata, outcome prices, best bid/ask. Cached 30–60s server-side and shared across visitors.
  • CLOB API: order books (refreshed every 10s) and price history.
  • Weather: NOAA/aviation station observations (METAR) plus Open-Meteo ensemble forecasts (ECMWF + GFS) at the settlement station.
  • Order books: Polymarket CLOB best bids/asks and top-of-book depth for NegRisk structure checks.

Current calculations

  • Market-implied probability is the midpoint of best bid and best ask, or the last listed price when there's no book.
  • Bracket distribution: implied probabilities normalized across mutually exclusive outcomes. The raw sum is shown so you can see the overround.
  • NegRisk buy-all worst case is $1 minus the sum of best asks across every outcome. It uses top of book only, with no fees or depth.
  • Weather forecast: each ensemble member gives a daily high/low for the station's local date; recent model-vs-station bias is removed, today's observed readings floor/cap the result, and members are pooled into whole-degree bracket probabilities.

Roadmap

  • 1. Public UI and real market discoverydone
  • 2. Internal trade-history research storedone
  • 3. Elon tweet forecast (XTracker counts, block-bootstrap model) vs marketdone
  • 4. Weather station forecasts (ensemble + observations)done
  • 5. Depth-aware NegRisk analyzer and conversion trackingplanned
  • 6. Walk-forward backtesting, Brier score, calibration curvesplanned

Limitations & disclaimers

  • Elon forecast: verified hourly counts from Polymarket XTracker; remaining posts are estimated by replaying real historical stretches of equal length starting at the same hour of day (recency-weighted, 14-day half-life) plus counting noise. It is not yet backtested or calibrated (milestone 6), so treat it as an estimate, not a validated prediction.
  • Historical trades of public reference accounts are studied internally as research. Nothing from them is applied to predictions until it passes backtesting.
  • Prices can be up to 60 seconds delayed. Order books can be up to 10 seconds delayed.
  • Read-only analytics. No trading, custody, or financial advice.