Soybean yield prediction in argentina using climate data

Basco, Emiliano; Elías, Diego; Aguirre, Maximiliano Gómez; Pastore, Luciana Working Paper Soybean yield prediction in argentina using climate data Economic Research Working Papers, No. 117 Provided in Cooperation with: Economic Research Department (ie), Central Bank of Argentina Suggested Citation: Basco, Emiliano; Elías, Diego; Aguirre, Maximiliano Gómez; Pastore, Luciana (2025) : Soybean yield prediction in argentina using climate data, Economic Research Working Papers, No. 117, Banco Central de la República Argentina (BCRA), Investigaciones Económicas (ie), Buenos Aires This Version is available at: https://hdl.handle.net/10419/322395 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-sa/4.0/ Working Paper N 117 | 2025 Soybean Yield Prediction in Argentina Using Climate Data Economic Research Working Papers | 2025 | N 117 Soybean Yield Prediction in Argentina Using Climate Data Emiliano Basco Banco Central de la República Argentina Diego Elías Banco Central de la República Argentina Maximiliano Gómez Aguirre Banco Central de la República Argentina Luciana Pastore Banco Central de la República Argentina May 2025 2 | BCRA | Working Papers 2017 Working Papers, N 117 Soybean Yield Prediction in Argentina Using Climate Data Emiliano Basco Banco Central de la República Argentina Diego Elías Banco Central de la República Argentina Maximiliano Gómez Aguirre Banco Central de la República Argentina Luciana Pastore Banco Central de la República Argentina May 2025 ISSN 1850-3977 Electronic Edition Reconquista 266, C1003ABF Ciudad Autónoma de Buenos Aires, Argentina Phone | 54 11 4348-3582 Email | [email protected] Website | www.bcra.gob.ar The opinions expressed in this paper are the sole responsibility of its authors and do not necessarily represent the position of the Central Bank of Argentina. The Working Papers series is comprised of preliminary material intended to stimulate academic debate and receive comments. This paper may not be referenced without authorization from the authors. 1 Soybean yield prediction in Argentina using climate data: A spatial panel and time series approach * Emiliano Basco†, Diego Elías†, Maximiliano Gómez Aguirre†, Luciana Pastore † May 2025 Abstract Agriculture, and especially soybean production, has a critical role in Argentina’s economy, as a major contributor to GDP and export revenue. This paper studies the impact of climate variability on soybean yields in Argentina using a novel department-level dataset spanning 1980–2023. We estimate a fixed effects spatial error model (SEM) to quantify the effects of weather shocks— measured by extreme heat, precipitation, and ENSO phases—while controlling for economic and technological factors such as seed technology and relative prices. Our results show that extreme heat significantly reduces yields, while moderate rainfall boosts them up to a nonlinear threshold. El Niño phases increase yields, whereas La Niña events are detrimental. Technological adoption and favorable price signals also enhance productivity. These findings highlight the importance of accounting for both climate dynamics and spatial distributions when estimating agricultural outcomes. Time series models provide a strong empirical basis for forecasting soybean yields and informing policy decisions under increasing climate uncertainty. These models can be employed as effective tools for anticipating yield outcomes under different climate scenarios and utilized in stress test exercises. This work provides valuable insights for policymaking decisions, contributing to prepare for potential economic impacts stemming from climate risks on Argentina’s agricultural sector. JEL Code: Q10, Q12, C13, C32, C33 Key words: Soybean Yields, Argentina, Forecasting, Model Selection * We are grateful to Tamma Carleton and Torsten Ehlers for their valuable comments and suggestions, which have greatly improved this work. We also thank Pablo Eduardo Abbate, Research Director at the Instituto Nacional de Tecnología Agropecuaria, and Cristian Russo, Chief of GEA at the Bolsa de Comercio de Rosario, for their insightful contributions and sectoral expertise. We appreciate the feedback received from participants at the 14th BIS CCA Research Conference, Bogota, 3 December and at the Research Department’s internal seminar at the Central Bank of Argentina (BCRA). The views expressed in this paper are those of the authors and do not necessarily reflect the views of the BCRA or its authorities. † Economic Research, BCRA. E-mail: e[email protected]ob.ar, [email protected], mauricio.gomezagu[email protected] and luciana.[email protected]b.ar 2 1. Introduction Agriculture is a fundamental sector within the Argentine economy. In 2021, primary agriculture and agrifood value chains contributed an estimated 16% to the nation’s GDP (World Bank, 2024). Agricultural exports account for around 60% of the country’s total exports, emphasizing the sector's importance in generating foreign currency and supporting economic stability. The soybean complex is the main agricultural production chain in Argentina, characterized by a strong export profile. It involves the industrialization of primary grain production and represents the country’s largest export chain, accounting for nearly 28% of total exports in 2022, surpassing both the cereal and automotive sectors. Soybeans are the second most important agricultural crop, following corn, averaging a production of 46 million tons annually over the last five campaigns (2018-2022). Another important aspect of the soybean sector is its volatility, since it experiences high production variability from year to year compared to other crops. Argentina plays a critical role in global agricultural markets, being the third largest producer and exporter of soybeans worldwide and the top exporter of soybean oil and soybean meal. Oilseed exports are a significant component of Argentina's trade, representing over 31% of total exports in recent years (e.g., 2021 and 2022). This sector's significant contribution reflects its importance for both domestic economic stability and global food markets. Climate events such as droughts in the Pampas Region have severe repercussions, decreasing yields of essential crops such as soybeans, corn, and wheat, and consequently affecting their entire supply chains. The impacts of these declines also influence multiple sectors within the economy. Reductions in agricultural production caused by droughts have a spillover effect, with implications for upstream and downstream industries. The agricultural sector's vulnerability to droughts also has broader economic consequences. A downturn in crop yields reduces fiscal revenue, lowers international reserves, and puts pressure on the exchange rate, impacting Argentina's economic resilience. For example, the recent drought in 2022-2023 had a significant impact on the yield and production of Argentina's main crops: wheat, corn, and soybeans. The effects were felt directly in agricultural production and, through the agribusiness chain, extended to other sectors. The substantial decline in exports also created a shock to government revenue and the balance of payments due to its influence on the exchange rate and international reserves. Two major climatic phenomena that impact moisture conditions and can cause droughts in the productive regions are La Niña and El Niño. La Niña results from cooling along the equatorial Pacific Ocean, leading to lower-than-average rainfall and causing droughts of varying severity across South American regions. In contrast, El Niño is a climate event marked by the warming of 3 the equatorial Pacific Ocean, which tends to bring above-average precipitation in Argentina. In this paper we consider the influence of El Niño and La Niña indicators, along with the rest of the climate information. In this paper, we examine the impact of climate variability on soybean yields in Argentina by combining a unique department level dataset covering 1980–2023 with daily weather observations—temperatures and precipitation—aligned to the crop’s phenological calendar. We augment these climatic controls with economic and technological variables, including international soybean and fertilizer prices, transgenic seed adoption rates, and land use change indicators. We employ spatial panel econometric techniques to capture both local fixed effects and spatially correlated unobservables, improving our understanding of how climate and regional interdependencies shape yield outcomes. Our empirical strategy unfolds in two stages. First, we estimate a fixed effects spatial error model (SEM) panel to quantify the long run relationship between weather shocks—measured by growing degree days, cumulative hours above 30 °C, precipitation, and El Niño and La Niña indicators—and soybean yields, while controlling for economic and technological covariates. We find that extreme heat consistently dampens yields, moderate precipitation increases them up to a nonlinear threshold, and lagged El Niño anomalies (warm phases) enhance yields while La Niña’s dry conditions are detrimental. Technological progress, as proxied by transgenic seed use and favorable price ratios, further elevates yields, and the highly significant spatial error term confirms the importance of unobserved, regionally correlated factors. In the second stage, we translate these insights into forecasting exercises by constructing different time series models based on the available weather information. We evaluate forecast performance via rolling window exercises and the Giacomini–White test, benchmarking against the USDA’s monthly yield forecasts with same data availability. We find early information models that systematically outperform USDA forecasts, while those models with information closer to the end of the growing season, our forecasts are as reliable as USDA ones. These findings carry important policy implications. By demonstrating that climate driven models can deliver accurate yield projections months before harvest, we provide a tool for anticipating foreign exchange pressures and informing monetary and fiscal policy responses. Our framework also lays the groundwork for constructing climate stress scenarios—simulating droughts, heat waves, or ENSO events—to evaluate potential economic vulnerabilities in Argentina’s agricultural sector. Finally, because the methodology relies on spatially disaggregated yield and weather data, it can be readily adapted to other crops or regions, underscoring the broader value of high resolution climate data in modern agricultural risk management and policy planning. 4 This paper contributes to the understanding in the relationship between climate variability and soybean yields in Argentina. We introduce a large data set that leverages georeferenced, delegation‑level records to uncover both local and regional climate–yield relationships and to account for spatial interdependencies among production zones. This new data set allows us to get a spatial error panel model that reveals how extreme heat, nonlinear precipitation, and ENSO phases affect soybean yields, while capturing technological gains. Finally, we provide statistically robust and accurate forecasts benchmarked against USDA projections. The paper is structured as follows: first, we present a literature review related to the influence of climate on agriculture. In Section 3, we present some concepts that are crucial to understanding soybean production in Argentina, along with our data sources, treatment of the data, and the calculations involved in constructing the dataset. Section 4 contains an overview of the econometric methodology employed, and we present the results of our analyses. The last part concludes and discusses the implications of our findings and recommendations for future research. 2. Related Literature A number of studies examine the effects of climate variations on crop yields. For example, Schlenker and Roberts (2009) studied the impact of variations of temperature on corn, soybeans, and cotton yields in the United States, using a panel data approach with nonlinear temperature effects. They found that extreme heat has significant negative impacts on yields, and defined heat thresholds for each crop, above which productivity is reduced. Deschênes and Greenstone (2007) explored the effects of annual variations in temperature and precipitations to estimate their influence on agricultural profits, using county level panel data in the United States. They estimated long-run effects and found heterogeneous and overall small outcomes across the US. Chen et al. (2013) conducted an analysis for corn and soybean yields in China, using spatial panel econometric techniques. They modeled the behavior of profit-maximizing farmers, and estimated changes on yields using climate variables, socioeconomic variables and variables representing farmers’ adaptation behaviors. They also found nonlinear and asymmetric relationships between yields and climate variables. Miao, Kanna and Huang (2016) examined the effects of climate variables and crop prices on corn and soybean yields and acreage in the United States, using a large spatial panel dataset. They found that prices have statistically significant effects on corn yields, and that the potential effect 5 of climate change on production is negative but highly heterogeneous, depending on climate scenarios and models. For Argentina, Cornejo and Ahumada (2021) analyzed the long-term relationships between climatic, technological, and economic factors and crop yields. They found that soybean yields adjust to technological improvements and that high temperatures have a negative effect. They also found evidence of CO2 fertilization. Cornejo (2021), using information on climate variables published in advance and with a frequency higher than soybean yields, found that there are forecasting gains when considering the maximum temperature during the plant's growth cycle. Additionally, when using precipitation data, the model based on annual data outperforms the others. These studies present different methodologies to assess the relationship between climate and agricultural productivity. Although technological advancements have improved overall yields, climate events, particularly droughts, continue to have a negative influence on agricultural production. 3. Descriptive Analysis and Data 3.1. Farmers production function Our study employs a production function as a framework to estimate the effects of climate conditions on yields. This method can be used to isolate the impact of weather on specific crops, while it is also necessary to capture the behavior of farmers and take into account for the range of compensatory responses they perform in reaction to climate variability. Soybean production, in particular, is determined by two key factors: the area planted and the yields achieved. The decision regarding the area to be cultivated is made in advance, based on expectations of profitability and risk. Producers take into account crop prices, macroeconomic conditions, and potential funding constraints. On the other hand, yield variability depends on multiple factors, including the application of inputs (fertilizers, pesticides), the adoption of advanced agricultural technologies, and the climatic conditions during the growing season. Price variables can also influence farmers' decisions on input use and production intensity. Our study focuses on yield variability as it is the variable which is more affected by climate shocks. While changing weather patterns, temperature and precipitation play a fundamental role in determining productivity, farmers decide the planted area before knowing the full extent of weather conditions. 12 Figure 6: average precipitation deviation from historical mean vs soybean yields 3.5. La Niña / El Niño Understanding the effects of El Niño and La Niña is crucial when analyzing Argentina’s climate conditions and their impact on soybean production. These phenomena significantly influence precipitation patterns in various regions worldwide, which, in turn, affect crop yields. Although these shifts can vary slightly between El Niño events, the most intense changes typically remain consistent across specific regions and seasons. The El Niño–Southern Oscillation (ENSO) is monitored through sea surface temperature (SST) anomalies in the equatorial Pacific Ocean. The primary measure used is the Oceanic Niño Index (ONI), which averages SST anomalies in the Niño 3.4 region (170°W–120°W). El Niño conditions occur when SST anomalies exceed 0.5°C for five consecutive months, indicating a warming phase. La Niña conditions are characterized by SST anomalies below – 0.5°C, signaling a cooling phase. These climate anomalies disrupt global rainfall patterns, affecting various agricultural regions worldwide. Unlike other soybean-producing regions, Argentina experiences reduced rainfall during La Niña years, leading to drought conditions that significantly impact crop yields. Historical observations confirm a negative correlation between La Niña events and soybean production in Argentina, as shown in Figure 7. This underscores the climatic vulnerability of Argentine agriculture to ENSO fluctuations. 13 Figure 7: Relationship between changes in soybean yield and El Niño/La Niña episodes We also explore whether early signals of La Niña events—typically available before the sowing season—have any measurable impact on soybean planting decisions, particularly in terms of total planted area. The evidence suggests that, in the Argentine context, anticipated La Niña conditions do not significantly affect the area sown with soybeans. As a result, the primary channel through which La Niña affects soybean production in Argentina is through its adverse effects on yields. This reinforces the relevance of focusing on yield responses—rather than acreage—as the key margin of adjustment when modeling climate impacts and building predictive models. 1 A key economic mechanism that could mitigate the impact of lower yields is price adjustment, which may dampen the broader macroeconomic effects. If La Niña–induced droughts were to significantly reduce global soybean supply, international prices could rise, partially offsetting farmers’ income losses. This, in turn, could help stabilize export revenues, consumption, and investment. However, this adjustment mechanism is less straightforward in the case of Argentina. The main reason is that La Niña does not affect all major soybean-producing regions equally. While Argentina tends to suffer from droughts and yield losses, Brazil and parts of the United 1 In meetings with specialists from the agricultural sector, they were specifically asked whether the risk of El Niño/La Niña events was taken into account in planting plans (such as the choice of crop mix, the more or less intensive use of inputs, etc.), as this could introduce some form of endogeneity. Their response was negative. 14 States often experience favorable growing conditions, leading to increased output in those regions (see Figure 8). As a result, global supply may not fall significantly, and price increases are limited. This is consistent with the evidence that the ENSO index does not systematically anticipate international soybean prices. Figure 8: El Niño and La Niña Teleconnections Map (Lenssen, Goddard & Mason, 2020) Figure 8.a. El Niño Figure 8.b. La Niña 15 3.6. Technology and other variables In addition to climate effects, it is necessary to consider the context of soybean cultivation in Argentina. This section describes some factors that allowed its expansion in volume and territory, which have led to the incorporation of specific variables in the models. Commercial soybean production began in Argentina in the 1970s. According to Cadenazzi (2009), soybean was initially introduced as a second crop after wheat, as its ease of management and adaptability represented a more profitable rotation, replacing the traditional agriculture/livestock rotation. In the 1980s, soil erosion effects were observed in the Pampas region due to intensive agricultural activities. However, by the mid-1990s, soybean production expanded with technological advancements that also allowed for the use of lower-quality land both in the core zone and outside of it, expanding the agricultural frontier and increasing yields. Internationally, a sustained increase in soybean demand was observed, driven by population growth and the need for food, as well as the growing demand for other uses. Better prices in the global market also stimulated supply. As mentioned earlier, before planting begins, producers will define the area to be cultivated, taking into account the market context and expectations. For this reason, the international price of soybeans and its recent volatility were incorporated as variables. The introduction of transgenic soybean in 1996, specifically RR soybean, marked a turning point in agricultural production in Argentina. This variety of soybeans is resistant to the broad-spectrum herbicide glyphosate, which eliminates all weeds. This translates into cost reductions as it is easy to apply. The practice of no-till farming also spread during this period. It involves eliminating plowing while the residue from the previous harvest conserves moisture and serves as fertilizer. Sowing is carried out with specially designed machines, with minimal soil disturbance. This method reduces fieldwork time and soil erosion. No-till farming, along with the use of RR soybean and glyphosate, complement each other since the former leads to an increase in weed quantity. These techniques, along with the development of new agrochemicals such as herbicides, pesticides, and fertilizers, as well as the development of specific machinery, contributed to the expansion of soybean production, as they increase yields and reduce production costs. This makes it feasible to produce in areas that were previously not viable, resulting in increases in both yields and cultivated areas. We incorporated the price and volatility of Diammonium Phosphate (DAP), a fertilizer commonly used for soybean production, as a signal of the variable costs for agricultural producers. 16 Additionally, a variable representing the proportion of the planted area where transgenic varieties were established was included as an indicator of the technological advancement in the production sector, as this represents a shift in the soybean production model. Information on technological advancement in each delegation was not available. With respect to land use, following Chen et al. (2013), we constructed an indicator of land use change, computing the year-on-year reduction in the planted area of crops other than soybean inside each delegation. This indicates the portion of cultivable land that can be made available to soybean production each year, and eventually have an effect on yields. 4. Methodological Approach and Results 4.1. Strategy for Model Selection In order to develop a tool with robust predictive capacity to predict soybean yields in Argentina, we explore different specifications using delegation-level data from across the entire country. These specifications differ based on the type of variable used, the frequency of the variables incorporated, the combination of frequencies (monthly, annual, and quarterly), and the construction of indicators from the same variable. This generates a multiplicity of data structures that can be used to predict yield, each of which requires different estimation methods. Our starting point is a panel data model designed to understand the underlying drivers of yield variation. While it is crucial to understand how climatic variables directly affect soybean yields in Argentina, these factors operate within an economic and financial context that can mitigate or exacerbate climatic effects. By detecting the most relevant variables in both the climate and the economic and financial aspects—and by examining their specific distributions—we are able to inform the configuration of alternative time series model specifications. 4.2. Panel Data Model: Specification and Results In this section, we present a detailed analysis of our panel regression methods and results. First, we test the panel unit root hypothesis, and then we proceed to estimate the proposed model using the fixed effects estimator for a stochastic production function. All panel unit root tests 2 —across multiple specifications—consistently reject the null hypothesis of a unit root, indicating that the series are stationary in levels and do not require differencing prior to estimation. Second, we estimate the proposed model using the fixed effects estimator for a stochastic production function. This approach relies in the production function mentioned in Section 3.1, along with production incentive variables, to achieve impact estimates by varying one or more 2 Levin, Lin & Chu (LLC) test, Im test, Pesaran & Shin (IPS) and Dickey Fuller 17 input variables, such as temperature, precipitation, the sea surface temperature indicator of the Pacific Ocean, and the use of incentive variables like soybean prices and input costs along with their volatility. We utilize the fixed effects model for our panel data for two reasons. The main reason is that the fixed effects model allows us to estimate a unit-specific effect for each delegation in the model. Additionally, the fixed effects model does not require the restrictive assumption that the specific effect of the delegation is independent of the included covariates, as is the case with the random effects model (See Appendix 2 for the Hausman test results). Also, we are using the total population (all the soybean producing delegations), which prevents us from needing a random effects estimation to generalize sample data to a broader population. The dependent variable for this model is the logarithm of soybean yields for each delegation (as described above). We estimate the following regression: log𝑌 𝑖𝑡 = 𝑍𝑖𝑡 𝛽 + 𝑐𝑖+ 𝜀𝑖𝑡 (1) 𝜀𝑖𝑡 = 𝜌 ∑𝑊𝑖,𝑗𝜀𝑗𝑡 + 𝜙𝑖𝑡𝑗 (2) Where log 𝑌 𝑖𝑡 indicates log crop yield in delegation i and year t. The term 𝑍𝑖𝑡𝛽 includes weather variables, a time trend and quadratic time trend, and other technological, land use change and economic variables. The term 𝑐𝑖 is included to account for the delegations’ fixed effects, and 𝜀𝑖𝑡 is the error term. Following Chen et al. (2013) and Schlenker et al. (2006), we allow for spatial correlation in the error term. In equation 2, 𝜌 is the spatial correlation factor, 𝑊𝑖,𝑗 is a spatial weighting matrix that identifies neighbors for each delegation, and 𝜙𝑖𝑡are the error terms that are independently normally distributed with E=0 and variance= 𝜎2. Panel models that include an interaction effect in the error term-called SEM (spatial error models)-, indicate that units might have a similar behavior because of shared unobserved characteristics (Elhorst, 2017). In this case, yield might be influenced by factors that are not included in the model, such as regional policies, soil quality or seed varieties used, that are spatially correlated over delegations. We test for the significance of the spatial error term and found it to be statistically significant (see Appendix 3). The model specification requires complete data for all delegations; consequently, we exclude seven delegations with missing or zero values from the estimation. The excluded delegations add up to 2% of the total production on average. 18 Table 2: SEM with spatial fixed effects. Coefficient Std. Error z P>|z| Trend -2,323 0,907 -2,560 0,010 Quadratic trend 0,001 0,000 2,570 0,010 El Niño temperature anomaly (-1) 0,030 0,006 4,590 0,000 Transgenic Soybean 0,328 0,087 3,760 0,000 Ratio of soybean price/fertilizer price (-1) 0,119 0,081 1,470 0,140 LUC indicator 0,087 0,051 1,700 0,088 GDD -0,000 0,000 -2,540 0,011 GDD over 30ºC -0,000 0,000 -2,280 0,023 Precipitation 0,002 0,000 7,530 0,000 Precipitation squared -0,000 0,000 -5,470 0,000 Spatial correlation 0,608 0,028 21,640 0,000 Number of groups = 32 Panel length = 35 The SEM panel estimation results are shown in Table 2, and the summary of the effects found is: • GDD and GDD over 30ºC: These variables represent the growing degree days and cumulated hours with temperatures above 30°C during the growing season respectively. Coefficients are both negative, small (due to scaling of the variables), and statistically significant. A negative coefficient associated with the variable GDD over 30ºC suggests that temperatures above 30 degrees negatively impact yields. The coefficient for the variable GDD is also negative, suggesting that the effect of high temperatures is dominant over the summer soybean growing season, having a negative effect on yields overall. • Precipitation and Precipitation squared: These variables reflect the cumulative precipitation and the squared cumulative precipitation between July and March along the same campaign. The coefficients associated with these variables (positive and negative, respectively) show a nonlinear relationship between precipitation and yields: more precipitation increases yields, but beyond a certain threshold, the impact starts to increase at a decreasing rate. • El Niño temperature anomaly: This variable reflects an expanded yearly Oceanic Niño Index (ONI), that shows the presence of the El Niño and La Niña events along with their intensity. The variable is lagged one year, to show the conditions prior and during the early stages of each campaign. The coefficient is positive and statistically significant, suggesting that warm conditions (associated with the El Niño events) are favorable for yield. Conversely, negative values of ONI indexes indicate La Niña events that promote 19 dry conditions over the soybean producing areas of Argentina, and would have a negative effect on yields. Regarding trend, market and technology variables, we incorporated the following: • Trend and Quadratic trend: The linear and quadratic trends represent long-term effects. The negative coefficient on the trend indicates a slight decrease in yield over time; while the positive quadratic trend suggests that the rate of change in yield may stabilize or even slightly increase in recent years due to factors not included in the estimation. • Ratio of soybean price/fertilizer price: It is the ratio between the soybean price variable and the price of DAP (Diammonium Phosphate fertilizer) on the year prior to sowing. The positive coefficient observed in this ratio indicates potential increases in farmer benefits, which in turn lead to higher yields. This is likely due to farmers trying to increase production per area in response to their perception of higher future benefits. • LUC indicator: The land use change indicator shows the year-on-year reduction in the planted area of crops other than soybean inside each delegation. The resulting coefficient is, although small, indicative of the fact that reductions in areas dedicated to other crops are being shifted to soybean planting. • Transgenic Soybean: The impact of the percentage of transgenic soybean use yields a positive and significant coefficient, indicating that as the use of transgenic soybean in the productive area has increased, it has contributed to improved yields regardless of climatic impacts. 4.3. Time series: model specification and data structure Using insights from the panel approach, we construct time‑series models to evaluate different strategies for predicting Argentina’s aggregate soybean yield. The independent variable remains the logarithm of soybean yields. For the explanatory variables, we generate a time series dataset based on the panel data, considering different aggregation methods. First, we average the delegations’ climate variables over the spatially interpolated data used in the panel model for the whole soybean areas. Second, we generate another dataset by averaging directly all soybean-producing zone’s meteorological stations records. Lastly, we create another dataset with climatic variables measured in the core zone only, in order to generate a leading indicator. The idea is to explore a uniform region in terms of climate, production, and management. We then test different combinations of variables. Our first set of time series models (Group G1) uses the same variables as the panel data estimation, where the weather variables were averaged 20 over the whole soybean-producing zone. In the second step, we identify the optimal model— based on entropy—using variables from the panel‐model for key month of the growing‐season calendar. This yields different time‐series models, each corresponding to the weather information available at successive stages (December, January, February, and March). To provide robustness to the comparison, we also incorporate a second group of models designed to capture different features of the phenomenon at hand. This second set of models (Group G2) contains combinations of variables different from those on group G1, and also takes into account the availability of data for different months. The explanatory variables are the total panel averages for all the soybean producing region. The third group (Group G3) takes into account different formulations using variables with annual, monthly, and quarterly frequencies, as is the case with ENSO. This mix of frequencies requires the use of different estimation methods such as MIDAS (Mixed Data Sampling 3 ). Models in group G3 use as explanatory variables the weather averages using indicators measured within the core zone only. The independent variable is the same as the previous groups, the logarithm of the total soybean yield in Argentina. 4.4. Forecast comparison Our model selection strategy is driven by the goal of anticipating or outperforming the forecast results published by the US Department of Agriculture (USDA), a globally recognized authority in agricultural production forecasting. Therefore, we consider two dimensions: the ability to anticipate and the ability to predict. In this context we use the term "anticipate" referring to the comparison between two models with information available at different points in time. A model with inputs that are more recent or closer in time to the event being predicted is expected to have greater predictive capacity than a model with inputs that are more distant in time. When a model with information that is more distant in time predicts a contemporary variable better than a model with more recent information, we say that the latter anticipates the former; even if these two models are not able to differentiate their predictive capacity (they predict the same thing). On the other hand, if two models’ inputs are available at the same time, the model that better predicts the target variable will be considered to have better forecasting power and therefore have better predictive capacity. To generate the forecasts, once the panel and time series models were defined, we divided the dataset into training and testing segments, a process often used in time-series analysis as setting 3 See Ghysels, Santa-Clara and Valkanov (2006) 21 an "estimation window." Within this window, each model is trained to make predictions, projecting ‘h’ steps ahead. For example, we estimate the model using data from 1989 to 2016 and we obtain a prediction for the soybean yields one year ahead. Then we expand the sample to 1989 to 2017, and we forecast 2018. Iterating this process results in a of RMSE for each model, to be compared with the USDA predictions’ RMSE. To compare the models, we use the Giacomini and White (2006) test, which allows us to evaluate, compare, and prioritize different models based on their predictive performance. This tool only indicates which model is the best within a set, but it does not provide information on how good the model is in absolute terms. To overcome this limitation, we use the USDA forecasts as a benchmark for comparison. Our benchmark is derived from a subset of the USDA’s series of predictions in their monthly reports. While USDA issues monthly Oilseeds reports, throughout the year, we take into account the reports published in January to April (the harvest starts in April and lasts until June). The reports are made available in the first days of each month, therefore we consider them to incorporate information up to the end of the previous month. For example, for the forecast included in the January report, we assume it incorporates information up until the last day of December of the previous year. Table 3 shows some of the results of the Giacomini-With tests. The test results allow us to compare all the forecasts coming from our panel Spatial Error Models (SEM) and Time series models from groups 1, 2 and 3 with published USDA forecasts. While we ran pairwise comparisons for all the available models, in this table we show the ones that outperform USDA forecasts 4 . Table 3: Model comparison (GW results) on models that perform better than USDA forecasts Model 1 Model 2 Data Availability Coeff. Std. Error t-Statistic Prob. SEM_DEC USDA_Jan_report DECEMBER -106,87 56,91 -1,87 0,1095 TS_G2_01_DEC USDA_Jan_report DECEMBER -115,53 65,51 -1,76 0,1283 TS_G3_03_JAN USDA_Feb_report JANUARY -150,24 80,26 1,87 0,1104 TS_G2_06_JAN USDA_Feb_report JANUARY -137,35 72,93 -1,88 0,1087 SEM_FEB USDA_Mar_report FEBRUARY -77,918 43,01 -1,81 0,1201 The first two columns show the models being compared. "Data Availability" refers to the moment when the data becomes available for input into the model. For example, in the first row, the SEM_DEC, -the spatial error modeland the USDA_Jan_report forecast take into account data 4 Results can be made available upon request to the authors. 28 Appendices Appendix 1: Causality tests When analyzing two lags, we find that the null hypothesis that prices do not cause production cannot be accepted; however, Argentine production has no effect on prices. Pairwise Granger Causality Tests Sample: 1980 2024 Lags: 2 Null Hypothesis: Obs F-Statistic Prob. DLPROD does not Granger Cause DLPRECIO 31 0.57428 0.5701 DLPRECIO does not Granger Cause DLPROD 3.57279 0.0426 The null hypothesis of no Granger causality from planted area to yields cannot be accepted; whereas, the hypothesis that yields do not Granger-cause planted area cannot be rejected. On the other hand, the null hypothesis of no Granger causality between planted area and ENSO cannot be rejected in either direction, indicating that the effect of ENSO on yields is isolated from any potential changes in planted area induced by ENSO. Pairwise Granger Causality Tests Sample: 1989 2022 Lags: 2 Null Hypothesis: Obs F-Statistic Prob. L_SUP does not Granger Cause ENSO 32 0.3726 0.6924 ENSO does not Granger Cause L_SUP 1.0931 0.3495 LOG_RINDE does not Granger Cause ENSO 32 2.6924 0.0859 ENSO does not Granger Cause LOG_RINDE 9.4367 0.0008 LOG_RINDE does not Granger Cause L_SUP 32 0.2634 0.7704 L_SUP does not Granger Cause LOG_RINDE 5.4377 0.0104 29 There is no evidence of bidirectional Granger causality between the international price and the ENSO index, indicating that they are statistically independent in this context. Pairwise Granger Causality Tests Sample: 1989 2022 Lags: 2 Null Hypothesis: Obs F-Statistic Prob. ENSO does not Granger Cause DLPRECIO 32 0.3471 0.7100 DLPRECIO does not Granger Cause ENSO 0.5429 0.5875 30 Appendix 2: Hausman test for Fixed Effects and Random Effects panels Since we have data from all relevant delegations (i.e., the sample covers the entire population of regions), panel fixed effects estimations are preferred because they capture the individual heterogeneity of each region without assuming that these differences are random. A random effects estimation would be more appropriate if the data were a random sample from a larger set of regions. To test the cross-section random effects, we conducted the Hausman Test, with the following results: Test Summary Chi-Sq. Statistic Chi-Sq. d.f. Prob. Cross-section random 45.666 10 1.648e-06 The results imply that fixed effects should be used. This suggests that the unobserved differences between regions are correlated with the explanatory variables, invalidating the key assumption of random effects. 31 Appendix 3: Moran and LM tests of spatial error correlation significance The spatial weight matrix used in the regression and for this test is a spatial contiguity matrix. This spatial contiguity matrix includes pairwise comparisons of all delegations, where the (i, j) element of W is unity if delegations i and j share a common boundary, and 0 otherwise. The matrix is normalized so that the sum of the elements in each row is equal to one. The Moran test is conducted over the cross-section of the delegations’ soybean yields for each year. The results are: Year Moran I p-value Year Moran I p-value Year Moran I p-value 1989 4,197 2,70E-05 2002 2,639 8,32E-03 2015 3,629 2,85E-04 1990 1,504 1,33E-01 2003 3,100 1,94E-03 2016 4,603 4,15E-06 1991 2,024 4,30E-02 2004 3,549 3,87E-04 2017 0,917 3,59E-01 1992 4,149 3,34E-05 2005 4,519 6,23E-06 2018 3,197 1,39E-03 1993 0,325 7,45E-01 2006 2,271 2,31E-02 2019 2,932 3,37E-03 1994 2,857 4,28E-03 2007 0,898 3,69E-01 2020 3,031 2,44E-03 1995 2,002 4,53E-02 2008 0,925 3,55E-01 2021 2,602 9,26E-03 1996 3,326 8,80E-04 2009 3,061 2,20E-03 2022 2,843 4,47E-03 1997 1,739 8,20E-02 2010 0,262 7,93E-01 1998 3,894 9,87E-05 2011 3,148 1,64E-03 1999 2,950 3,18E-03 2012 6,473 9,62E-11 2000 3,612 3,03E-04 2013 4,096 4,21E-05 2001 0,232 8,17E-01 2014 0,659 5,10E-01 32 Appendix 4: SEM Panel models for forecasting We develop an alternative variable for all SEM forecasting models (SEM_DEC, SEM_JAN, SEM_FEB and SEM_MAR) that aims to capture within-year temperature patterns more effectively, which may provide valuable insights for yield prediction. To show the procedure applied, the following table describes the SEM_MAR model. The main difference between the Spatial Error Model (SEM) in Table 2 and the SEM_ MAR model lies in how they incorporate the frequency of the El Niño temperature anomaly. Specifically, when we refer to the El Niño temperature anomaly (-1) in the SEM Table 2 model, we are referring to the average temperature during the calendar year that starts just before the campaign. In this context, the anomaly (-1) includes temperature information both prior to and during the campaign. While this annual average anomaly provides useful information about the overall impact of temperature on yield, it may also obscure important intra-annual patterns that are relevant for forecasting. Coefficient Std. Error z P>|z| Trend -2,748 0,951 -2,890 0,004 Quadratic trend 0,001 0,000 2,900 0,004 El Niño temperature anomaly (Dec(-1)) 0,178 0,044 4,070 - El Niño temperature anomaly (Mar(-1)) -0,180 0,060 -2,990 0,003 El Niño temperature anomaly (June(-1)) 0,057 0,038 1,500 0,134 Transgenic Soybean 0,312 0,095 3,290 0,001 Ratio of soybean price/fertilizer price (-1) 0,173 0,084 2,060 0,040 LUC indicator 0,084 0,051 1,660 0,097 GDD -0,000 0,000 -2,620 0,009 GDD over 30ºC -0,000 0,000 -2,260 0,024 Precipitation 0,002 0,000 7,560 - Precipitation squared -0,000 0,000 -5,570 - Spatial correlation 0,595 0,029 20,580 0,000 Number of groups = 32 Panel length = 35 In the table above we show the quarterly pattern, where we incorporate the quarterly variables that end in December, March and June, previous to the start of the campaign. The difference in signs between temperature in different periods indicates that different combinations of quarterly temperature could give different results of the soybean yields even if the average is the same. 33 Appendix 5: Estimation results: SEM with spatial fixed effects SEM December SEM January SEM February SEM March Trend -2,060 -2,192* -3,084*** -2,748*** (1,257) (1,152) (1,017) (0,951) Quadratic trend 0,001* 0,001* 0,001*** 0,001*** (0,000) (0,000) (0,000) (0,000) El Niño temperature anomaly (-3Q) 0,219*** 0,220*** 0,202*** 0,178*** (0,059) (0,054) (0,047) (0,044) El Niño temperature anomaly (-2Q) -0,261*** -0,252*** -0,213*** -0,180*** (0,080) (0,073) (0,065) (0,060) El Niño temperature anomaly (-1Q) 0,139*** 0,113** 0,068 0,057 (0,051) (0,047) (0,042) (0,038) Transgenic Soybean 0,347*** 0,339*** 0,372*** 0,312*** (0,130) (0,118) (0,102) (0,095) Ratio of soybean price/fertilizer price (-1) 0,257** 0,194* 0,196** 0,173** (0,112) (0,104) (0,090) (0,084) LUC indicator 0,077 0,076 0,079 0,084* (0,052) (0,052) (0,051) (0,051) Growing degree days -0,0004* -0,0003** -0,0003*** -0,0002*** (0,000) (0,000) (0,000) (0,000) GDD over 30ºC -0,000058 -0,000007 -0,000139 -0,000260** (0,000) (0,000) (0,000) (0,000) Precipitation 0,001*** 0,002*** 0,002*** 0,002*** (0,000) (0,000) (0,000) (0,000) Precipitation squared - 0,000001*** -0,000001*** -0,000001*** -0,000001*** (0,000) (0,000) (0,000) (0,000) Spatial correlation 0,694*** 0,668*** 0,624*** 0,595*** (0,023) (0,025) (0,028) (0,029) Number of groups = 32 Panel length = 35 Standard errors are reported in parentheses. *, **, *** indicates significance at the 90%, 95%, and 99% level, respectively