US natural gas storage forecast

Model vs Trading Economics survey consensus vs EIA actual · updated

Phase 4 gate
—
Week endingModelWN3ConsensusActual Model errCons errSurpriseSource
Paper trades (QG, Wed close → Thu close)
WeekStatusSiden Pred surpriseActual surprise EntryExitP&LReasoning

Consensus is logged independently of the model and never used as a feature. Street surveys (Reuters/NGI) are paywalled; the automated series is the Trading Economics economist survey, ingested on GitHub Actions. EIA actuals overlay the official Thursday 10:30 ET WNGSR print (ir.eia.gov/ngs/wngsr.json) so a lagging Open Data API cannot leave the latest week blank. The gate requires 26 overlapping weeks with rolling MAE below consensus and surprise-direction hit rate above 50%. Paper trading is a QG overlay on predicted surprise versus consensus; it does not wait for the gate. WeatherNext 3 is an A/B forecast source (purple) and does not drive paper trades. Next-week predictions require 7 complete forecast days — Maps Weather API is 10 days, so week+1 is usually available from Wednesday; full WeatherNext 3 (15 days) covers it from Saturday. Rules are in natgas/signal/translate.py. Ledger: data/paper/trades.csv.