Figure 3.

Results of applying CA to the country-product matrix. (a): The country-product incidence matrix shows a triangular structure. Column and rows are sorted by the first CA axes, known as ECI and PCI for the country and the product similarity matrices, respectively. (b): Sorted spectrum for the country-product matrix (log-scale). The slow decay and lack of clear gaps in the spectrum suggests a high-dimensional, homogeneous dataset. Numerical labels are reported for the first four eigenvalues, and for the twenty-first and twenty-second (the small gap between the latter two motivated the choice of a twenty-dimensional embedding in the final part of the analysis) (c): Correspondence plot showing the first (horizontal) and second (vertical) CA axes for the country similarity matrix. The first CA axis is known as the ECI, and explains $3.5\%$ for the total variation. The second axis explains $2.5\%$ of total variation and seems to distinguish countries specializing in garments and textiles from other countries. Colors indicate the obtained clusters when running k-Means with $K=3$ on the embedding spanned by the first 20 CA axes. (d): GDP per capita as a function of the first CA axis (ECI). The dashed line is the linear regression of log(GDPpc) with ECI ($R^2$=0.49). Colors indicate the same clusters as in c). The analysis was performed using our SCCA R package²¹.