Co-operation between Rochdi FEKI and Nouri CHTOUROU began when the latter ascertained that many of the relationships within economics are non-linear and that there was a need to establish models which the dynamic aspects of these relationships. Occurrences of non-linearity had been encountered during research by CHTOUROU into Governance and Economic Development. Indeed, his goal was to seek verification through appropriate linkages between the institutional structure of a country and its levels of economic and social performance. From then appeared two obviousnesses: for each country, institutional retroactive compartments between them and their effects on growth follow nonlinear transmission ways. Furthermore, since we cannot treat a phenomenon that cannot be subject to measurement (intangibles) he also addressed the difficulty of how to measure the quality of governance within an institutional setting. This latter (intangible) quality that governance includes multidimensional layers indicated that any attempt to measure it required using some form of aggregation technique to construct a composite indicator.
FEKI’s solution was to propose using Artificial Neural Network (ANN) models which contained proven qualities in modelling complex relationships. The major advantage of neural networks is their ability to learn dependencies between variables based on a finite number of observations. This feature makes them flexible enough to model complex relationships without needing any a priori assumptions about the distribution of variables (a major constraint of conventional statistical techniques).
Initial mathematical results have shown that ANNs produce excellent nonlinear models, particularly through the features of the sparse approximation they have. In addition, the development of algorithmic techniques for rapid and accurate learning, allowed neural networks to increasingly establish themselves as the methodology of choice in fields such as robotics, pattern recognition, signal analysis, medical diagnosis, financial, etc..
There are a large number of artificial neural networks, the most popular being multilayer perceptrons and self organizing maps developed and introduced by Teuvo Kohonen in the 80s. Kohonen’s neural model is a data visualization technique, (according to a fixed topology), for projecting an input space into a multidimensional output space, usually one-dimensional or two dimensional. This projection is such that those elements having similar characteristics are grouped in the same class of a previously defined map. Thus, Kohonen maps perform two tasks: the first is the reduction in size, while the second is to reveal similarities. In the context of data analysis, this technique appears to be a particularly useful method of clustering. Individuals are grouped into classes according the topology of the input space. This means that we define a priori the notion of neighbourhood between classes, ensuring that the elements contained in a neighbourhood area (within the same class or related classes) have similar characteristics.
The application of these self-organizing (Kohonen) maps to analyze the problem of macroeconomic governance initiated the collaboration and subsequent research between CHTOUROU and FEKI. Significantly, these early uses for the purpose of analysis and typological clustering generated sufficient positive results and evidence leading to the development of a new method of constructing composite indicators and rankings. This ambitious programme has taken several years of research and led to the design of CFAR-m (Chtourou Feki Aggregation and Ranking method).