Geospatial Weather Features Definitions

Descartes Labs price models use external geospatial features to enhance their forecasting power. Some of these features rely on commonly used time-series data in the commodities industry, however others are directly derived from unique data that Descartes Labs generates. This article will detail some of these geospatial features and how they are specifically engineered in a way industry practitioners may use them. 

Naming Conventions

Each of these features share the following common naming convention, which can be useful to interpret their contributions to a model forecast:

wxsupply_gen2_{location}_{weather-variable}_{transformation}_{crop-period}

Location

For each unique geography, we use locations that are relevant to a particular crop or set of crops. For example, in the Robusta Coffee model we use:

• Vietnam
• Indonesia
• Brazil - Rondonia
• Brazil - Espirito Santo

Weather Variables

We also use different weather statistics, such as:

• t2mean: Mean temperature for that day in Celsius
• t2min: Minimum temperature for that day in Celsius
• tpcum: Cumulatied Precipitation for the day in meters

Transformations

We also apply various transformations to the time-series observations, including:

• none: No transformation applied
• norm: 10 year normals
• deviation: Deviation from the 10 year normals
• deviation_accum: Accumulated deviation over a period

Crop Period

For each commodity's growing period, we implement period-specific date ranges. For example, in the Robusta Coffee model in Brazil we will not indicate the same period as Vietnam:

• Flowering Period
• Cherry Growth Period
• Harvest Period