Production Forecasting Models

A technical summary of commodity production forecasts at Descartes Labs.

Description

The Market Insights production forecasts use geospatial weather features to predict the end-of-season production figures.

Weather Features

The various features used as inputs into prediction models are as follows:

  • Evapotranspiration
  • Vapor pressure deficit
  • Solar radiation
  • Accumulated precipitation anomalies
  • Frost index
  • Primary modes of tropical variability:
    • Madden Julian Oscillation (MJO)
    • El Niño Southern Oscillation (ENSO)

Forecast Components

Each forecast has two components. Note that the end-of-season is fixed, so the forecast timeframe shortens as the calendar year progresses. 

  1. 0-30 days uses the Global Ensemble Forecast System forecast model 00z run
  2. 30+ days uses a proprietary seasonal forecast model that forecasts through the end of the season

Miscellaneous Details on the Forecast Models

  • All models are trained on production values obtained from Conab for Brazil and the USDA in the US from 2008 to the present. 
  • Each model is trained every 5th day of the growing season. 
    • For example, Coffee production in Brazil is from August through July.
  • Back testing has been performed over the past 10 years using a leave-one-out approach, where the model is trained on all years except one.