An adaptation of a prototype recursive model for evaluating and predicting micro-macro economic data in Botswana

Abstract

Naturally arising patterns in data are known to present potential sources of useful information at both micro and macro-economic levels. We carry out unsupervised and supervised modelling of Botswana's macro-economic data attributes obtained from disparate sources. Both techniques, commonly used to detect inherent patterns in data, have adjustable parameters which inevitably vary across applications. Thus, we propose a sequential unsupervised-supervised modelling approach in which Exploratory Data Analysis (EDA) is used to detect basic structures in data which are then passed on an algorithm based on the Expectation-Maximisation (EM) mechanics. The EM convergent values are then used to guide data labelling before applying the neural networks model. We demonstrate how future economic structures may be detected, monitored and managed by iteratively focusing on conditional checks of a generic algorithm. For the purposes of modelling robustness, we propose setting up an integrated data repository and source that would provide data-based guidelines to policy makers in addressing the country's economic issues while providing economic researchers access to data and/or information resources. Outstanding issues are identified and discussed and potential future directions are clearly highlighted. Key words: Data mining, data over-fitting, data recycling, EM algorithm, neural networks, knowledge extraction from data, supervised learning, unsupervised learning.