Know more

At AgroParisTech, 16 rue Claude Bernard à Paris, at 13h 30, amphi Coleou, Rotem Zelingher will defend her thesis : "Agricultural Commodity Price Forecasting Using Comprehensive Machine-Learning Techniques".

Abstract:

Would it be possible to develop a forecasting tool for agricultural commodity prices that is both accurate and interpretable, and above all, accessible to the general public? Thanks to such a tool, the forecasting and analysis of food prices could cease being a domain exclusively reserved to scientists, economists and financiers, and turn into an implementable instrument used by whoever is concerned by food security. Such a tool would allow those who do not have the financial capacity or the appropriate technical background to forecast agricultural commodity prices at least a month in advance. Ideally, this tool can be updated
sequentially, so that its results could evolve with the market.

This PhD explores the feasibility of this idea in three parts.

The objective of the first part is to test the ability of several statistical and machine learning models to simulate changes in maize prices based on annual changes in maize production and yield observed in major producing regions. Two approaches are examined: quantitative, allowing to simulate price variations continuously, and binary classification, to enable calculating the probability of price increases or decreases. In this first part, the models are fitted into the data, evaluated using statistical criteria, and finally, identify regions that strongly influence price variations. The results show that, relative to the last quarter of the year, maize production in North America has, by far, the highest impact on global crop prices. They also demonstrate the possibility to establish a relationship between production variations in North America and variations in the world price of maize.

The second part of the thesis applies the models developed in the first part and adapt them to produce monthly forecasts of maize prices. We compare the performance of these models to that of forecasting techniques often used for time series analysis. We show that, for short-term forecasts usually up to three months, time series analysis techniques have several advantages. First, they are generally more accurate; second, they allow the performance of direct price change forecasting without using any additional information. On the other hand, to forecast prices at longer time horizons, we would usually prefer machine learning models, as they yield results of lower error. These algorithms have an additional advantage: they can diagnose probable causes of extreme price variations and identify which production shocks cause these variations.

Finally, in the third part, we extend the one-crop model to consider two other very different crops - soybeans (a legume that constitutes the world-main source of vegetative protein) and cocoa (a crop produced in only a few countries located in the tropics). We evaluate the forecasting ability of the techniques developed in the previous stages to predict price changes for soybeans and cocoa. Additionally, we test the sensitivity of the results relative to three geographic scales: regional, national and continental. Also demonstrated in this chapter is the application of machine learning methods to identify which production shocks drive price shocks. Overall, this thesis shows that machine learning methods are a potential tool for understanding and forecasting the impact of agricultural production on price variations. These approaches can be easily implemented since they rely on publicly available data, accessible via public website. These tools can thus contribute to democratising the analysis and forecasting of variation in agricultural commodity prices.