Grain & Oilseed Yield - Leveraging Data to Improve Predictions
We were approached by a client in the agricultural grain origination and farm services industry who wanted to determine the impact of adverse weather on grain production in their geographical business area. The client had a thesis that consistent adverse weather had a negative impact on critical times during the life cycle of the commodities they handle but needed data-based analysis to validate this.
The Execution
We applied a machine learning strategy to gather key production data for corn, soybeans, and wheat in 6 prominent grain producing states in the northern and western corn/soybean and western/northern wheat belts in the United States. We then extracted acreage, prevent plant acres, yield, and production data in the major producing counties in each state. Additionally, we obtained county-level precipitation, high-low-mean temperatures, growing degree days, and palmer index data, from which we developed a proprietary machine learning model that allowed us to create an index of "adverse weather events" in each county.
Using multiple proprietary machine learning models, such as Random Forest, Gradient Boosting, and Neural Network models, we evaluated historical "adverse weather events" and production data over a 15-year period to determine if there were any regional weather-production irregularities. After several iterations and fine-tuning the models, we identified a model that resulted in a high confidence evaluation of the data analyzed.
We combined the client's market expertise with our data mining, data architecture, analysis, and machine learning approach. This resulted in a well-informed and confident senior management team ready to evaluate alternative strategies to address their business. This includes mitigating exposure to adverse weather, using machine learning to optimize data to their economic advantage, expanding the use of business intelligence to unlock analysis and creativity in the organization to strengthen its role as a market leader further, and evaluating tens of millions of rows of data. The final model leveraged historical weather patterns and adverse conditions through growing seasons to predict final grain production at the county level.