Using Self-Organizing Maps to Identify Operational Risk

Abstract

In recent years, the awareness and concern for operational risk in financial
institutions have increased, and several disastrous events in the last two
decades been caused by human error. With this, the regulatory demands
have increased on the financial institutions to control operational risk.
One operational risk that Svenska Handelsbanken AB (SHB) has detected
are the audit changes of trades which occur when a trade need some form of
altering from its original state, which can lead to losses for the bank. The
bank has looked into identifying and forecasting these losses with the use of
a neural network clustering method called Self-Organizing Map.
This thesis expands on a previous project initialized by SHB on the po-
tential of using this method to identify operational risk, and research the
robustness and effectiveness of the Self-Organizing Map and trying to ob-
tain an optimal solution by using quantifiable measurements like Matthew’s
Correlation Coefficient.
By evaluating the algorithm through visualizations of the generated maps
and evaluating its prediction ability through Cross-Validation, the results
obtained from this thesis indicate that the Self-Organizing Map has great
potential in this area and is able to identify these risks with a relatively high
accuracy.

Popular Abstract (Undetermined)

The current need and requirement for financial
institutions to manage their risks have never
been higher and new ways to handle and
forecast these risks are in constant development
in these institutions. The machine learning
method of Self-Organizing Map introduced in
the 1980’s by Teuvo Kohonen proposes the use
of a neural network and big data analysis to
cluster data, and research has now been
conducted to see if it can predict operational
risk in the form of audit changed trades.
Self-Organizing Map is an unsupervised machine
learning method where input data of a high
dimension organizes itself in a lower-dimensional
grid. The map can consist of for example a
rectangular grid and a hexagonal grid of neurons in
two dimensions, and the input data can be of
unlimited dimensions. The sought out results from
this method is to capture non-observed latent
processes in the input data and find clusters within
the map. The latent processes proposed in this
article is that the operational risk of audit changed
trades can cluster itself in the map after training. An
audit changed trade is a trade where the input has in
some way been altered from its initial state, or even
cancelled, which can lead to losses if severe enough.
The process is to first select the features that every
trade will consist of, features that hopefully can have
an impact on the human error that causes a risky
audit changed trade, e.g. the trader, the type of
change, the price, the nominal amount etc.. The
number of features then determines the dimension
in the input data. The input data consists of a
number of trades labeled as risky, and a number of
unlabeled trades. The neurons in the map then,
besides being assigned its coordinates in the map,
get assigned the same number of randomly
generated weights as the number of dimensions in
the input data. While iterating over an initially
determined number of epochs, the neuron’s weights
get updated to become more similar to the input
data. One data point in the input data gets randomly
selected and the neuron in the map with the most
similar weights gets selected as the best matching
unit. This neuron, and every neuron within its
neighborhood, determined by a certain
measurement, gets updated according to a selected
adoption rate to become more similar to the
selected data point. To finish of the map after the
last iteration, every data point finds its best
matching unit and gets placed there in the map, see
Figure 1.
To identify a cluster within the map, the best ratio
between the True Positive rate, the number of
identify labeled trades divided by the total number
of labeled trades, and the False Positive rate, the
number of falsely identified unlabeled trades by the
total number of unlabeled trades, are determined by
expanding the area around the cluster's center which
is considered as the identified cluster. By selecting
the neurons in the cluster which have at least one
labeled trade as the area where a trade is risky, the
prediction ability of the map can be determined.
By using the data set provided and the described
method, the Self-Organizing Map in the project that
this article is based on was able to detect a risky
trade with a 74 % accuracy and a non-risky trade
with almost 100 %. This shows that the use of a
Self-Organizing Map to identify operational risk has
great potential if implemented in a financial
institution and a viable source for the task of
identifying risky audit changed trades.

• author: Ljungblom, Lukas and Berggren, Jonatan
• supervisor: Erik Lindström
• organization: Mathematical Statistics
• course: FMS820 20181
• year: 2018
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Self-organizing Map, Machine Learning, Neural Network, Clustering, Operational Risk, Audit Change
• language: English
• id: 8948029
• date added to LUP: 2018-06-11 15:34:47
• date last changed: 2018-06-14 07:34:54