Recipient Organization
TEXAS A&M UNIVERSITY
750 AGRONOMY RD STE 2701
COLLEGE STATION,TX 77843-0001
Performing Department
Agri Economics
Non Technical Summary
The project seeks to evaluate the implied errors or shocks to the future "history" of economic variables from a set of policy restrictions. That is to say, the project will compare the implied path from a set of agricultural policies with the historical observed errros in order to assess how plausible or likely it is that the implied paths could ever occur. Based on historical measures of theory supported price, quantity and instrumental variables observed over recent time, we fit a dynamic model and project forward months and years into the future the size of the errors on all series in the dynamic model. These errors are then compared with already known errors from previous years and months to assess the likely realization of the particultura policy. If large errors are required, the policy analyst might reasses the magnitude of the policy restriction or may offer additional evidence supporting why the relatively large errors are likely to be forthcoming.
Animal Health Component
80%
Research Effort Categories
Basic
20%
Applied
80%
Developmental
(N/A)
Goals / Objectives
The overall objective is to build and apply Machine Learning Time Series (MLVARs) models capable of generating conditional policy paths on agricultural economic systems. The prime objective is to articulate the mathematics of forecasting conditional policy paths (on a subset of VAR variables) and the implied dynamic effect on other variables in the VAR. Applications are seen as generating ex ante paths (forecasts) in areas of agricultural systems under policy restrictions (e.g., agricultural export restrictions, fuels use restrictions, or food price restrictions). With respect to conflict and poverty, projections of policy paths under different food price ceilings in conflict prone zones (following Bessler et al., 2016), will allow us to state quantitatively the implied path on conflict and fatalities from conflict. Here the objective is to model the stochastic difference equation between conflict, poverty, fatalities, and agricultural crop prices and project (ex ante) the consequences (and associated probabilities) of particular ceilings on food prices (or other policy restrictions). We have begun initial work on this problem in the Democratic Republic of Congo. Here programs for conflict alleviation requires a long term commitment to a set of policy conditionals (troop levels or crop prices over time). Other applications, including import restrictions related to food safety (see Costa, et al., (2015) and Huang, et al., (2015)) and food insecurity should be forthcoming, once the underlying mathematics and programming are completed.
Project Methods
The mathematical model underlying the project is the vector autoregression (Sims 1980), as applied many times by the principle investigator (Bessler (1984). For any application we need a sufficiently rich set of time series data (where theory and problem setting dictate what in particular is to be the focus). Innovations from the estimated VAR are studied with machine learning algorithms (Spirtes, Glymour and Scheines (2000)), to identify data-based structure. This last step originated here at TAMU in Bessler and Akleman (1998) and has been replicated by many others over the years (Demiralp and Hoover (2003), Hoover (2005), Moneta et al., (2013), Duangnate and Mjelde (2017), among others). Of particular recent interest are advances in this work on machine learning under non-Gaussian innovations (errors from the VAR). Economic and financial data are notoriously non-Gaussian. Recent algorithm development with non-Gaussian data has led to an improved ability to identify structure from the underlying data (see the applications in (Moneta et al., (2013) and Huang, Lai and Bessler (2017)). Based on the fit model outlined here, one then projects forward any distance into the future of the underlying VAR. Here, one holds fixed the future path on the policy variable (or variables), calculating the implied effect on the other innovation (error) terms (other series) required to generated the policy path with least distance (size of the implied errors). The original mathematics for this conditional forecasting problem has been around for more than thirty years (Sims (1982)). However, that work assumed a priori knowledge of contemporaneous structure. Our innovation (contribution) will be to use Machine Learning to allow a data generated structure as the initial driver of the conditional forecasts. [It is certainly a worthy research question to ask whether the a prioi structure generates "better" forecasts relative to the data generated structure. This questions will be addressed.] The (general) software for doing this is programmed in ESTIMA Doan (2008). Our initial task will be to interface these programs with the machine learning models to offer a unified model for analyzing conditional policy studies. The results from our implied policy path (ex ante) can then be compared (after the fact) to the actual path taken by all of the variables in the underlying VAR.