Source: TEXAS A&M UNIVERSITY submitted to
MACHINE LEARNING AND ECONOMETRICS FOR AGRICULTURE POLICY
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
TERMINATED
Funding Source
HATCH
Reporting Frequency
Annual
Accession No.
0221098
Grant No.
(N/A)
Project No.
TEX09354
Proposal No.
(N/A)
Multistate No.
(N/A)
Program Code
(N/A)
Project Start Date
Dec 22, 2009
Project End Date
Dec 21, 2014
Grant Year
(N/A)
Project Director
Bessler, DA, A.
Recipient Organization
TEXAS A&M UNIVERSITY
750 AGRONOMY RD STE 2701
COLLEGE STATION,TX 77843-0001
Performing Department
Agri Economics
Non Technical Summary
For much of the twentieth century econometrics, applied with observational data, proceeded strongly from an a priori model. Specification was based on an explicit model identified using zero restrictions based on a priori theory. The resulting model, sometimes labeled "structural" was generally grounded in the ceteris paribus assumption to provide over-identifying restrictions to disentangle correlations from underlying causal structures. In the last two decades of the twentieth century researchers began a serious questioning of the strong use of a priori theory to suggest identifying restrictions on observational economic data (Sims 1980). In particular the constructed models did not forecasts new data points well. Rausser (1982) offered the following assessment of the state of affairs in agriculture in the late 1970's: To the U.S. government officials who were struggling to control inflation. the tremendous increase in food prices was indeed a bitter disappointment. At this juncture, it became crystal clear that the constructed models of the USDA were no longer viable. The forecasts generated by these models appeared to be outliers in comparison to the actual behavior of the system. (page 2) The model that replaced the earlier over-identified structural models in much applied work was a time series representation of just identified models. These models generally forecasted new data points better than the previously used structural models; see (Nelson 1972), Granger and Newbold (1986) and Bessler and Brandt (1981). Early versions of this time series representation were devoid of explicit theory, other than that used to select the variables to be studied in a multiple time series representation. However, in order to use these models for policy purposes, research workers had to specify a structural ordering, at least, on contemporaneous relationships among variables studied (Bernanke 1986). For the most part such specifications were subjective, based on priors of the researchers. New ideas on inductive inference, emanating from computer science and philosophy in the late 1980's and in the 1990's, have given research workers in economics a less subjective, more objective, base from which to inform time series models with respect to contemporaneous structure. The structural vector autoregression or structural error correction model introduced in Swanson and Granger (1997) and Bessler and Akleman (1998) gave researchers a more solid footing for using the structural VAR in policy (see Awokuse and Bessler (2003) and Hoover (2005)). These new ideas in inductive inference are generally discussed under the rubric of "machine learning" and are given formal treatment in Spirtes, Glymour and Scheines (2000) and Pearl (2009). These methods have contributed to a rich literature in applied econometrics over the last decade, but have yet to make inroads into policy-making via conditional forecasts (see Sims (1982) for an early view of policy, time series models and forecasts).
Animal Health Component
(N/A)
Research Effort Categories
Basic
(N/A)
Applied
(N/A)
Developmental
(N/A)
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
6097310301050%
6037310301020%
6031719301010%
6033399301010%
6026110301010%
Knowledge Area
602 - Business Management, Finance, and Taxation; 603 - Market Economics; 609 - Economic Theory and Methods;

Subject Of Investigation
6110 - Economy of the United States and sectors thereof; 7310 - Experimental design and statistical methods; 1719 - Cotton, other; 3399 - Beef cattle, general/other;

Field Of Science
3010 - Economics;
Goals / Objectives
The objective of this project is to study further agricultural economic policy based on econometric specifications generated via advances in machine learning and time series econometric methods. Specifically the work will involve constructing large-scale time series econometric models that are capable of generating conditional forecasts of futures values of important variables in the US agricultural economy. Methods for combining a large number of series for inclusion in the times series models will be a major focus of the effort (see the factor VAR methods discussed below). Outputs will be contributions to the refereed literature demonstrating the conditional forecasting ability of the constructed policy models on US agricultural products.
Project Methods
Building on earlier work on US cattle markets (Bessler and Kling (1989), Bessler and Davis (2004), Henry, et al. (1992), and Stockton, Bessler and Wilson (2009)), wheat markets (Bessler and Babula (1987)) and West African millet markets (Bessler and Krenga (2002) and Vitale and Bessler (2006)), this project will focus on building large-scale policy models. The work will again be guided by economic theory for choosing categories of variables; however, instead of focusing on a small set of variables (as the above mentioned papers do) we will study further how to combine dozens of potentially relevant variables into a time series policy model. Our initial investigations will consider statistical dimension reduction methods such as factor and principal component analyses, which treat theoretical constructs as unobserved factors. Stock and Watson (2002), Bernanke, Boivin, and Eliasz (2005) and Kwon (2007) give us the factor-augmented vector autoregressive model (FAVAR) based on the idea that a large number of variables are required to provide us adequate information about the economy (the agricultural economy). Such information can be effectively incorporated in a model by a small number of estimated factors. The standard VAR augmented with estimated factors is a natural way to embrace large information sets. Our use of machine learning will provide us with the ability (which Stock and Watson (2002) and Bernanke, Boivin and Eliasz (2005) do not exploit) to offer empirical evidence on contemporaneous structure on innovations from the factor VAR. This last step gives us the ability to make policy projections that are consistent with the causal structures found in the data (Kwon 2007).

Progress 12/22/09 to 12/21/14

Outputs
Target Audience: Economists and policy-makers interested in the dynamics of agricultural prices. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided? We advised several PhD students who used the techniques developed under this project. Further the work developed here found its way into both undergraduate and graduate teaching of the project leader. How have the results been disseminated to communities of interest? Lectures at national (American Agricultural Economics Association 2009, 2010, 2011, 2012 and 2013; Carnegie Mellon University 2013, Southern Agriocultural Economics Association 2013 and 2014 ) and internatinal (MASH University of Paris Sorbonne May 2014) conferences. Over fifteen papers, based on work emanating from this project, were published in the refereeed literature (listed elsewhere in this report). What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? We demonstrated that macjhine learning techinques can be used with well-known econometric techniques to give researchers a better view of the dynamic relationships holding prices and quantities together in domestic (US) and international economies. Key to this accomplishment was our ability to idetify causal structures in contemporaneous time. These results were demonstrated in both small systems (five to ten variables) as well as large systems (over 100 variables).

Publications


    Progress 10/01/13 to 09/30/14

    Outputs
    Target Audience: We appied machine learning ideas and algorithms to offer a clear idea of: price transmission in cattle markets in Mali; soft drink price transmission and quantity sold for two major soft drink lables in Chicago, USA; and evidence of financial contagion among Mercosur nations in South America. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided? One PhD student completed her study under my co-direction: Shiva, Layla PhD 2014, Texas A and M University, Economic Analysis of Voluntary Offset Market and Bioenergy Policies. How have the results been disseminated to communities of interest? The journal articles listed above are one way I have disseminated my work. In addition I lectured on machine learning at the University of Paris in May 2014: Econometrics, Causality, and Machine Learning,Modeles et Apprentissage en Sciences Humanities et Sociales, University de Paris, Pantheon-Sorbonne, May 22-23 2014. What do you plan to do during the next reporting period to accomplish the goals? I work with three PhD student who will complete their study with Factor Vector Autoregression and machine learning within the next year or two (timing is not clear here).

    Impacts
    What was accomplished under these goals? We provided researchers with new applications of machine learning in economic systems. We demonstrated, as well, the ability of these algorithms to exploit the non-Gaussian nature of the underlying data to uncover causal relations in data.

    Publications

    • Type: Journal Articles Status: Awaiting Publication Year Published: 2014 Citation: Olsen, K.K., J.W. Mjelde, and D.A. Bessler, 2014. Price Formulation and the Law of One Price in Internationally Linked Markets: An Examination of the Natural Gas Markets in the U.S. and Canada. The Annals of Regional Science.
    • Type: Journal Articles Status: Awaiting Publication Year Published: 2014 Citation: Lai, Pei-Chun and D.A. Bessler, 2014. Price Discovery between Carbonated Soft Drink Manufactures and Retailers: A Disaggregate Analysis with PC and LiNGAM Algorithms, Journal of Applied Economics.
    • Type: Journal Articles Status: Awaiting Publication Year Published: 2014 Citation: Bizimana, J.C., J.P. Angerer, D.A. Bessler and F. Keita, 2014. Cattle Markets Integration and Price Discovery: The Case of Mali, Journal of Development Studies.
    • Type: Journal Articles Status: Awaiting Publication Year Published: 2014 Citation: Viale, A.M., D.A. Bessler and J.W. Kolari. 2014. On the Structure of Financial Contagion: Econometric Tests and Mercosur Evidence, Journal of Applied Economics, 17:373-400.


    Progress 01/01/13 to 09/30/13

    Outputs
    Target Audience: Professionals interested in the use of econometric models for decision-making. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided? Nothing Reported How have the results been disseminated to communities of interest? The professional literature (Journal of Agricultural and Applied Economics and Journal of Applied Economics) has been used as the primary source of communication. Further a presentation by the project leader (Bessler) at Carnegie Mellon University: On MicroEconomics: The Use of TETRAD for Model Specification, was delivered at CMU in Pittsburgh in October of 2013. What do you plan to do during the next reporting period to accomplish the goals? I will explore further the use of factor models to reduce the size of the data set studied.

    Impacts
    What was accomplished under these goals? We articulated conditions for valid inference in econometric models using observational data. In particular we focused on "what can we say when such inference is undertaken in the presence of latent variables?" Applications are offered with respect to macroeconomic analysis of US business failures and nutrient policies in US fruit and vegetable trade.

    Publications

    • Type: Journal Articles Status: Published Year Published: 2013 Citation: Palma, M. L. Ribera, and D.A. Bessler. 2013. Implications of U.S. Trade Agreements and U.S. Nutrition Policies for Produce Production, Demand, and Trade. Journal of Agricultural and Applied Economics 43(3):465-480
    • Type: Journal Articles Status: Published Year Published: 2013 Citation: Zhang, J., D.A. Bessler and D.J. Leatham, 2013. Aggregate Business Failures and Macroeconomic Conditions: A VAR Look at the US Between 1980 and 2004. Journal of Applied Economics 16:183-208
    • Type: Journal Articles Status: Published Year Published: 2013 Citation: Bessler, D.A. 2013. On Agricultural Econometrics, Journal of Agricultural and Applied Economics 45(3):341-48


    Progress 01/01/12 to 12/31/12

    Outputs
    OUTPUTS: Ideas on conditional independence from machine learning have been communicated to economists and social scientists via academic publications, dissertations and project proposals. We offer help in understanding causal mechanisms in settings where theories are weak or non-existant. PARTICIPANTS: Not relevant to this project. TARGET AUDIENCES: Nothing significant to report during this reporting period. PROJECT MODIFICATIONS: Nothing significant to report during this reporting period.

    Impacts
    Our major accomplishment was showing the conditional independence among forecasts from several althenative models or theories imply the direction of information flow amongst those models or theories and consequently a preference by users.

    Publications

    • Bessler, D.A. and Z. Wang, 2012. D-Separation, Forecasting, and Economic Science: A Conjecture, Theory and Decision, 295-314.
    • Jean-Claude Bizimana, PhD 2012. Texas A&M University. Essays on Dynamics of Cattle Prices in Three Developing Countries of Mali, Kenya and Tanzania.
    • Sung Wook Hong PhD 2012, Texas A&M University, Three Essays on Price Dynamics and Causation among Energy Markets and Macroeconomic Information.


    Progress 01/01/11 to 12/31/11

    Outputs
    OUTPUTS: Ideas considered and developed under this project have been used to sort-out underlying causal structures in the areas of the economics of bio-security and international equity market inter-relations under alternative monetary exchange rate aggregations. Linking these machine learning ideas to earlier work in econometrics has been a major output under the project. PARTICIPANTS: Not relevant to this project. TARGET AUDIENCES: Academics reading the econometric literature have been the target of this project. Three publications have been forthcoming during this reporting period. PROJECT MODIFICATIONS: Not relevant to this project.

    Impacts
    The work under this project and related earlier projects has been cited over 40 times in the professional literature (as counted in SCI).

    Publications

    • Attavanich, Witsanu, B.A.McCarl, and D.A. Bessler, 2011.The Effect of H1N1 (Swine Flu) Media Coverage on Agricultural Commodity Markets,Applied Economic Perspectives and Policy 33:241-59.
    • Bessler, D.A., J.W. Kolari and T. Maung. 2011. Dynamic Linkages Among Equity Markets: Local Versus Basket Currencies, Applied Economics 43:1703-1719.
    • Kwon, Dae-Heum and D.A. Bessler, 2011.Graphical Methods, Inductive Causal Inference, and Econometrics: A Literature Review, Computational Economics 38:85-106.


    Progress 01/01/10 to 12/31/10

    Outputs
    OUTPUTS: Fundamental ideas and procedures for machine learning in economics and agricultural policy have been explored and applied to areas of food safety, bio-security, and price discovery. Such work has been communicated via refereed literature and conference proceedings. PARTICIPANTS: Nothing significant to report during this reporting period. TARGET AUDIENCES: Nothing significant to report during this reporting period. PROJECT MODIFICATIONS: Nothing significant to report during this reporting period.

    Impacts
    Ideas discussed and applied in this research have been cited over thirty times this last year in the professional literature. These cites are to current research as well as similar work actually published in years prior to 2010, but all relating to the thrust of the current project. The key idea is machine learning ideas studied by the PI are finding applications in agricultural economics and economics.

    Publications

    • Stockton, M.C., D.A. Bessler Roger K. Wilson, 2010. Price Discovery in Nebraska Cash Cattle Markets, Journal of Agricultural and Applied Economics 42(1):1-14.
    • Palma, M.A., Y. Chen, C. Hall, D.A. Bessler, and D. Leatham. 2010. Consumer Preferences for Potted Orchids in the Hawaiian Market, HortTechnology 20(1):239-244.
    • Chong, H., M. Zey and D.A. Bessler. 2010. On Corporate Structure, Strategy and Performance: A Study with Directed Acyclic Graphs and PC Algorithm, Managerial and Decision Economics 31:47-62.