Forecasting is an indispensable tool in the decision making process. People, on a daily basis, make forecasts inherently and decide on the basis of that. Formal forecasting methods are commonly applied to substantiate business decision making. Agriculture is no exception to this. There are numerous examples of agricultural projections on production, consumption, prices and trade. A South-African example of these types of projections is the outlook generated by the Bureau for Food and Agricultural Policy (BFAP). The formal forecasting scene in South Africa is, however, predominated by models that generate annual forecasts, with structural models. This is incidentally also the method used by the BFAP research program, mentioned above.
Logically, more frequent forecasts lead to better decision making. Since formal high(er) frequency forecasts are absent from the South African milieu, this study aims to determine whether the generation of these type of forecasts are feasible and meaningful within the South African agricultural context. Thus, a forecasting evaluation is applied to average monthly yellow maize prices in South Africa. The variable modelled was developed by taking the average monthly price of daily yellow maize closing prices as reported by the Agricultural Marketing Division (AMD) of the South African Futures Exchange of the Johannesburg Stock Exchange (JSE). This series is considered due to the strategic role that yellow maize plays in the South African Agricultural industry.
Several methods were considered as possibilities to be used in generating monthly projections. The study was, however, steered into time series econometrics arena because of data availability and the initial objectives of the study. Relevant variables are mostly available in a quarterly or yearly frequency. This problem is even more severe in macro-economic variables, such as gross domestic product (GDP) and disposable income, which are also used in agricultural modelling. Practical considerations guided the study to models which are parsimonious and easily updated. Time series models adhere to these qualities, resulting in two time series methods being applied. Firstly, univariate models, which generate forecasts based on the stochastic properties of the underlying data generating process, are regarded. Secondly a multivariate model is estimated to determine whether additional fundamental information improves on univariate models. Both univariate and multivariate models produce unrealistic forecasts, since both models forecast a mean level with little or no variation from month to month. If a 12 month forecasting period is considered, the monthly forecast, with no or little variability around the mean is essentially the same as an annual forecast. This ads little value to short(er) term decision making.
The study attempts to understand why the time series method yielded unrealistic results, since it has been applied successfully in the following studies Gjolberg (1997) and Liew, Shittan and Hussain (2000), as far as univariate models are concerned, and Skaggs and Snyders (1992) and Colino (2008) in the case of multivariate models. Possible causes for the mentioned unrealistic results projected with the univariate model are: Non-linear underlying data generating process, volatility clustering in the data generating process and non-normal distributions of average monthly yellow maize prices. Upon further inspection it is apparent that these are common characteristics of financial time series, which caused the study to venture into the financial econometric field. These characteristics were not initially considered, since it was not the objective to make forecasts to facilitate investment decisions in financial markets, but rather to serve as a tool for grass root decision makers in the agricultural industry. Multivariate models fared poorly in forecasting as well, since there were already problems in the cointegration establishment stage of model development. A general expectation in the market is that world maize prices and the Rand/Dollar exchange rates should be exogenous variables with respect to South African yellow maize prices and therefore, cointegration could only be established with parity prices. This confirms that world maize prices and Rand/Dollar exchange rates are only drivers of local maize prices when South Africa is trading under an import or export parity regime and prices are formed close to or at parity prices. This, however, has little benefit for forecasting.
A theory that goes hand in hand with the concept of commodity forecasting is that of the efficient market hypothesis (EMH). Although it was not initially the objective to test this hypothesis, discussing the effect of an associated futures market impact on price forecasting necessitated a referral to this theory. A question that beckoned was whether time series methods were simply the wrong method to use when attempting to generate monthly yellow maize price projections, or is forecasting of this price series a futile exercise? The latter was found to be true in that the market was determined to be weakly efficient. There are however still forecasting methods that could be applied even if the EMH holds. These however falls under advanced topics of forecasting and were not included in the scope of the study.
Formal high frequency forecasts in agriculture are to a large extent absent from South African agricultural literature and even more so in agricultural economics literature. This study attempts to address this by testing various methods. However none of the methods that were applied were successful in forecasting monthly yellow maize prices. The possible reasons for the performance lack of the models are explained in detail and a number of interesting conclusions regarding the usefulness of forecasting models for agricultural future markets are made. Apart from many empirical issues, one of the reasons stems from the fact that forecasting is a subject matter that stands with legs in different, and frequently far removed, study fields. This is even more true for forecasting prices of a commodity with an associated derivative market.
This study however provides a synthesis of various subject matters. Themes that are usually conducted in isolation, such as market efficiency, finical econometrics, agricultural practicalities and price formation, were combined into one study. The need for more diligent record keeping is illustrated by the empirical findings and the importance of exposing agricultural economics scholars to a broader scope of econometric methods and concepts is identified. In the South African context this study is, important, since it serves as a starting point for future research in agricultural forecasting in general, specifically for scholars wanting to explore the “higher” frequency milieu. Forecasting research opportunities that might be explored in the future are inter alia, neural networks, bootstrapping techniques and the application of time series methods (as applied in this study) for products or commodities with a different price formation process.
© 2010, University of Pretoria. All rights reserved. The copyright in this work vests in the University of Pretoria. No part of this work may be reproduced or transmitted in any form or by any means, without the prior written permission of the University of Pretoria.
Please cite as follows:
Labuschagne, MH 2010, Testing alternative methods for forecasting maize prices in South Africa, MCom dissertation, University of Pretoria, Pretoria, viewed yymmdd < http://upetd.up.ac.za/thesis/available/etd-07202011-123322/ >