Page 98 - Invited Paper Session (IPS) - Volume 2

IPS184 Celestino G. et al.
imbalances across the economy. For instance, the same external deficit might
be consistent with deficits in the public or in the private sectors and, in the
latter case, with deficits for households or for corporations. Similarly, an
external surplus might be accompanied with relative large deficits for some
domestic sectors compensated by even larger surpluses for others. More in
general, the distribution by sector and size of imbalances might present very
different, complex configurations that might have a strong influence on the
effects of shocks in the economy, and that make insufficient a country
characterisation merely on the basis of the overall external deficit or surplus.
We propose a country characterisation and grouping criteria that take into
account such differences. Given the potentially large complexity in the
distribution of sector imbalances, any classification approach that relies only
on human analysis would run the risk of suffering from “confirmation bias” or
other cognitive biases. To partially correct this, we make use of a deep learning
technique, Self-Organizing Map (SOM), a specific kind of neural network that
2
identifies patterns in complex data.
This approach can be applied to a much larger set of indicators, the
analysis in this note being a “proof of concept” for such usage at a larger scale.
However, interpreting the results in the way that is done across the note would
not be possible with such larger indicator sets. To help bridge that difficulty
we propose applying SOM to data previously transformed via Principal
Component Analysis (PCA) which allows reducing the dimensionality of the
outcome (although at the expense of decreasing the direct interpretability of
the components/ indicators examined) . Even though we work here with a
reduced set of indicators, we also apply PCA to it in a second exercise to
provide a geometric interpretation of the SOM results.
The next section of this paper discusses briefly the applied methodology. The
third section shows the data analysis. The last section summarises and briefly
discusses avenues for future work.

2. Methodology
Our data set consists of quarterly time series for the net lending/net
borrowing of the individual domestic sectors - non-financial corporations
(NFCs), financial corporations, general government and households (including
non-profit institutions serving households)- for all countries in the euro area.
We work with medians of four quarter rolling sums expressed in percentage
of GDP. The medians are calculated for the ranges1999-2008 and 2009-2018,
i.e. for the pre-crisis and crisis and post-crisis period, when the distribution of
sector imbalances showed large changes. We then have 38 countryperiod

Similar approaches have been applied by: López Iturriaga and Pastor 2013. Peltonen and Sarlin
2
2011.
85 | I S I W S C 2 0 1 9