How VanderSat satellite data is improving cocoa trading strategy

Commodity trading, Agriculture

This case study highlights the results of one of our cocoa yield prediction projects. Let us show you how our scientific progress and unique high resolution satellite data can impact your trading strategy.

 

Unsurprisingly, commodity traders are more than eager to understand how they can harness the power of these Ag datasets to benefit their business. Insight and information are the only things capable of enhancing – or indeed hindering – your ability to make a profit in a market. Finding ways to create more useful insights and generate more information than other market players is exactly what VanderSat can do for commodity traders all over the world.

 

The data VanderSat delivers on climatic factors such as moisture, vegetation and temperature can go a long way towards predicting the final crop yield. From tropical zones to temperate zones and even irrigated land, our data provides the latest crop information. In addition to daily monitoring of conditions, we are able to supply historical data for any location on Earth. This historical data – going back to 2002 – provides an understanding of how today’s values relate to growing seasons almost two decades ago. This enables VanderSat not only to supply essential data but also to assess the impact of the situation you see today.

 

To understand how these changing conditions impact crop and final yield, we compare the historical VanderSat soil moisture dataset to historical production data. For one cocoa trader, we used this knowledge to develop and provide a national cocoa forecast for Ivory Coast and Ghana. We were able to deliver an absolute monthly yield forecast up to six months ahead, with a correlation of (R2) > 0.9. This improved on a previous model fed solely by pod-counting data.

 

VanderSat began the project by processing the satellite soil moisture data for Ghana and Ivory Coast (2002-present, see Figure 1) and researching the physical behaviour of the specific crop and its reaction to a range of climatic conditions. Does drought always have a negative impact? Can cold night-time temperatures prevent disease? Does the crop thrive when soil moisture is above a certain threshold? These initial questions helped produce a final model based on hardcore science, infinitely more reliable than the traditional ‘black box’ approach.

Adding the VanderSat data to the model resulted in an increase in forecast accuracy of up to 20%

Figure 1. Time series showing mean soil moisture for 2015-04 to 2019-12 for the central eastern region of Ivory Coast. The blue points indicate measurements, the red line shows the 20 day average and the black line shows the normal climatic conditions.

Once we are satisfied with our knowledge of a crop and the quality of the time series based on our data for the region (Figure 1), we turn our attention to delivering the end result. A dataset or a prediction has to be integrated into the client’s current trading forecasting strategies or it will never achieve its full potential.

 

To elevate the client’s position with the final deliverable, we need to discuss exactly what type of product we want to supply and how we plan to deliver it. That’s why we usually offer two different paths: either VanderSat develops a model to supply forecasts as a service, or we create and upkeep a parameter dataset which allows our client’s in-house scientists to improve the results of their own models simply and efficiently.

 

After studying the crop and the region’s soil moisture and temperature data intensively, VanderSat then correlates our soil moisture data with the final crop output in the region of interest. For cocoa, final output can be divided into two crops: the mid-crop which grows during the dry season, and the main crop which grows during the wet season. Yield data was available for four full growing seasons, so mid-crop and main-crop output could be compared to soil moisture anomalies in the most impactful months as determined by preliminary research.

 

The graphs in Figure 2 show how soil moisture anomalies relate to crop production: the September/October anomaly in relation to main crop production (top left) and production of both crops (bottom left), and the December/February anomaly in relation to mid-crop production (top right) and production of both crops (bottom right). The months selected were those when soil moisture and other climatic conditions are considered most detrimental or beneficial to crop outcomes. September and October relate directly to the survival of the main crop, which is generally harvested in November, while December is important for fruit setting, and February for the survival of the mid-crop fruits. A lower soil moisture anomaly indicates a lower final production and a higher soil moisture anomaly a higher final production level.

 

The physical background knowledge and a clear-cut correlation between our climatic variables and the final prediction led to the creation of a robust forecast.

Figure 2. Graphs showing the soil moisture anomaly on the x-axis and production of the specified crop on the y-axis. Each red point indicates one year of production in relation to the soil moisture anomaly in certain months; they show a clear correlation

A final model is created to predict the agreed target, in this case absolute national yield on a monthly basis. The target is predicted using a broad but specific set of parameters created by VanderSat based on our in-house expertise and the correlations found. These parameters include soil moisture deviation from the climatic mean, consecutive dry days, cumulative temperature anomalies, total days with soil moisture below a selected threshold, and the moment when the crop surpasses a set biomass level. The time invested in creating specific parameters for a specific crop and region enables us to optimize the model’s predictive power based on previously found correlations in order to deliver a highly robust model. The final model’s accuracy and performance is then tested by getting it to predict an unfamiliar dataset, in this case the last year of data (Figure 3, based on an out-of-sample method). The monthly predictions for this project are delivered up to six months ahead and are updated yearly.

Figure 3. Graph comparing real production (red line) to (out-of-fold) predicted production 1,2,3 and 4 months prior to harvest for the 2018/19 growing season. The production is shown as x 1000 tonnes.

Figure 4. The four prediction times on the x-axis show the forecast improvement expressed as a percentage (in SI,  the error relative to the range) compared to the base model without VanderSat climatic data. The biggest improvement is for two months (20%) and the smallest for four months.

Let me know how we can improve your commodity trading strategies

Let’s talk

Dr. Robin van der Schalie
Senior Remote Sensing Scientistrvanderschalie@vandersat.com

Related cases