Learning joint probabilistic weather forecasts from station observations alone

Key points

• CLARA (Calibrated Advection-Routing Attention) learns joint Gaussian predictive distributions of five surface variables from station observations alone, without numerical weather prediction or reanalysis; the approximately 28,000-parameter model supports CPU training and prediction.1
• Holding marginal variances fixed, removing learned correlations worsens joint negative log-likelihood by 1.0-2.8 nats per station.1
• A covariance-scale estimator, proved consistent under stated assumptions, improves short-lead calibration but over-corrects at long leads.1
• Synthetic interventions show an attention-bias coefficient alone does not measure forecast influence.1

Sources (1)

• [1]Learning joint probabilistic weather forecasts from station observations alone
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 11:57 AM

  Assessing compound weather risks requires forecasts representing dependence between variables.

  CLARA (Calibrated Advection-Routing Attention) learns joint Gaussian predictive distributions of five surface variables from station observations alone, without numerical weather prediction or reanalysis; the approximately 28,000-parameter model supports CPU training and prediction.

Extractive summary: sentences quoted from the sources.