3.1 Rain and Supply
A large number of firms produce soybeans in Argentina, Brazil, and the United States and sell their output in a competitive global market. Between 1996 and 2010, these three countries had close to an market share, as seen in Table I. Deodhar and Sheldon (1997) studied the degree of competition in the soymeal export market using data with annual frequency from 1966 to 1993 and found that the market was very close to perfectly competitive. Beans are grown during the extended summer season lasting from the beginning of May to the end of October in the United States, and from the beginning of November to the end of April in Latin America.
This is a model for a single growing season (6 months long). Let T be the harvesting time at the end of the current growing season, measured in days. Because Latin America and the United States are in opposite hemispheres, only one of these two regions is actively growing soybeans. For the active region, total local supply at T,
as seen from is uncertain because production will depend on weather events.
The supply from inactive regions at T is the inventory that remains from earlier harvests (the last one being roughly 6 months in the past). This is inventory available in April in the United States, or inventory available in Latin America in October. Figure 1 represents the chronology of the main events. Consider, to fix ideas, the growing season from month 0 to month 6 in Figure 1. In this case the active region is the United States, while Latin America is inactive. I assume that the rate of soybean consumption within a growing season is known; therefore, inventory at the inactive region is predictable as seen from . In particular, the ratio of existing stocks in the United States, Brazil, and Argentina at the beginning of the local year to subsequent yearly production for the period 1997–2010 was most often less than 10%, and never more than 20% (tables 21–23 in the Oilseeds Report of December 2010, USDA). Taking into account this small 10% initial inventory prior to the last harvest, a plausible linear consumption rule implies that supply at the inactive region roughly 6 months after the last harvest is
where is the long-term average regional yearly output. Total global supply available at T is
Hence, total relative supply available at T, as a fraction of yearly global production, was relatively more stable than the individual outputs of the United States or Argentina over the same period. This is due to fact that the United States and Latin America had comparable outputs in this period, and the mitigating effect of inventories. It is important to note that output from Brazil is appropriately included in the present analysis for the accounting of the level of global supply. However, in this paper I will focus on supply shocks in Argentina and the US Midwest, for reasons already discussed in Section 'THE GLOBAL SOYBEAN MARKET'.
Uncertainty about supply at harvesting time is a function of the yield per unit of land, which depends on technology and water intake. In the absence of irrigation, water intake is provided exclusively by rain. Let be the amount of rain fallen on day j over a region of interest and
the total precipitation during the midsummer period. Based on the positive relationship between precipitation and yields documented by Tannura et al. (2008) I postulate a linear relationship between midsummer rain and terminal output at harvesting time
where is the long-term historical yield. Expectations of terminal output are updated during the midsummer reflecting changes in expected rain precipitation. Taking conditional expectations in (7) I write
It is learning about rain at the prediction horizon that drives the daily change in expectations in (11). In a more realistic setting, weather forecasting skill is less than perfect and certainty about rain on a given target date develops gradually as time evolves toward the target date. Bickel et al. (2011) evaluated the accuracy of Probability of Precipitation (PoP) forecasts generated by the National Weather Service dependent on NOAA, the Weather Channel, and CustomWeather, which are available to market participants at low or no cost. They considered forecasts at 753 locations in the United States and found evidence of significant forecasting skill for up to a 96-hour horizon for all weather forecast providers. This leads me to postulate
From (9) and (12) it follows that the impact of rain in expected supply is
for a rescaled error term . An alternative to the approach described above consists in using actual weather forecasts rather than subsequent precipitation. However, there is no publicly available record of historical weather forecasts for Argentina going back as far as 1996. Regarding US weather forecasts, Chincarini (2011) reports that archived forecasts for the National Weather Service Model Output Statistics begin in September 2005.
3.2 Supply and Prices
World supply in (15) is the sum of supply at T in active and inactive regions. The latter contribute to total supply but are assumed to be certain because they are remaining inventories from earlier harvests. As such, the day-to-day variation in expected supply at T is driven by changes in expected supply in the active regions. Taking conditional expectations with respect to the information available on consecutive days in the growing season, summing over active regions, and using (13) I obtain
Next, let . I assume that changes in price within a growing season are relatively small so that
Next, take conditional expectation with respect to on (15) and use (18) to obtain
Also, taking conditional expectation in (18) with respect to information on consecutive days leads to
which, by (4) is
where is the market share of the region of interest in the global production of soybeans and dt is a deterministic trend.
Then, (22) can be expressed using (23) to obtain
This is a highly stylized model for a fully integrated global market in which an additional unit of supply anywhere in the world has equal impact on the CME price traded in Chicago. Supply shocks, in turn, depend on regional rain, market shares, and physical characteristics of each region.
The first quantitative prediction of the model (24) is that excess local rain precipitation leads to a decrease in the CME future price through the combination of the price elasticity of demand and the regional yield sensitivity with respect to rain .
In this model, rain in the active region has a positive impact on supply at T, and therefore a negative impact on the price for delivery then. As represented in Figure 1, the same consumption/inventory holding rule in (2), implies that
The third prediction of the model (24) is that the sensitivity of the CME price with respect to regional rain is proportional to . This is a testable implication that follows directly from the assumption about global integration of supply. A special case of interest is that of physically similar regions such as the US Midwest and the central region of Argentina. If both regions have equal sensitivity , then the response of the CME price to local rain is simply proportional to the local share of the global market. Rather intuitively, according to this model the CME price should be more responsive to weather events on those regions that produce larger quantities of soybeans.
The model assumes that a single variable measuring total rain over an entire region predicts regional output. This is clearly a simplification since rain and yields also exhibit variability within each region. However, attempts to identify the price impact of rain on smaller regions (e.g., individual states) are likely to be hampered by high rain correlation between neighboring states and the lesser economic importance of each individual state. The partition in two very distinct regions adopted in this paper seems appropriate to investigate the implications of globally distributed production.
The model is silent about other factors that might affect soybean prices. Tannura et al. (2008) showed that soybean output is sensitive to summer temperatures; therefore, I add temperature fluctuations as a control in the estimation stage. Prices are also sensitive to demand fluctuations and financial speculation. These issues, although certainly very important in the determination of commodity prices, seem at first order unrelated to the main goal of this paper, namely evaluating the impact of the distributed nature of production. Some of these effects, formally included in the noise term in this Section, are taken into account in the empirical estimation of the model through appropriate controls.