Page 253 - Contributed Paper Session (CPS) - Volume 3

CPS2011 Dominique H. et al.
therefore particularly interesting to explore the dynamics of firms from this
sector.
Most of the literature focuses on the financial services or the information
technology sectors. Hence, this paper adds value to the sector analysis also.
Further, the paper also considers a longer time-window of 11 years from 2006-
16. This serves two purposes: it encompasses the sub-prime period and at least
one complete economic cycle.
The findings from the analyses bring out interesting perspectives. For
example, they show how Amazon differs from its peers in the sector, how
Macy’s can be differentiated from others on certain financial aspects, etc. Such
insights can help investors understand these firms better and make their
investment decisions accordingly.

2. Methodology
The two-step approach adopted in this paper includes a dimension
reduction of the indicators, i.e., the financial ratios, and using the dimension
scores as input variables for non-linear clustering analyses using SOM.
SOM is a type of competitive neural network that projects a high-
dimensional input space on prototypes of a low-dimensional regular grid. It
accomplishes two goals: to reduce dimensions and to display similarities
among prototypes. The algorithm aims at clustering together similar
observations while preserving the original topology of the data (i.e., similar
observations in the input space are clustered together into the same unit or
into neighbouring units on the map) (Olteanu, M., & Villa-Vialaneix, N, 2015).
Specifically, in contrast with other artificial neural networks which apply error
correction learning, SOM is based on competitive learning, where the output
nodes compete among themselves to produce the winning node (or neuron).
Only the winning node and its neighbourhoods are activated by a particular
input observation. This architecture allows SOM to preserve the topological
properties of the input space. Therefore, SOM can be effectively utilized to
visualize and explore the properties of the data.
SOM is a two-layer feed forward and completely connected network as
shown in Figure 1. The data from the input layer are passed along directly to
the output layer. The output layer is represented in the form of a grid, usually
in one or two dimensions, and typically in the shape of a rectangle or hexagon.
The algorithm can be summarized in the following 3 steps:
1. Initialization: initialize the neurons’ weights;
2. Competing: ①compute the scoring function (such as Euclidean distance)
for each output node; ② locate the winning node (the closest match with
the input)
3. Learning: update the weight vectors of the winning node and its
neighbours using a linear combination of the input vector () and the

242 | I S I W S C 2 0 1 9