Potentials - November/December 2017 - 10

This shift to a small number of assumptions
(assumption-free approaches are possible)
by statistical inference is, perhaps,
the element most difficult and troubling
for many people to grasp.
of data scientists to conduct the ML
training and evaluation.
Numerous classification problems
have been solved by ML. Many of the
most successful ones fit the supervised
learning domain, whereby the training
data contain the correct classification
results. There is considerable research
on identifying objects, such as whether
the object crossing the street in front
of a self-driving car is a human, a
deer, or another automobile (Fig. 1).
Clearly, this type of classification may
have life-threatening consequences.
Facial recognition is gaining higher
accuracy, and language translation
is in a similar situation, with fast
recent improvements being made.
These breakthroughs are generating expectations for the field, and
most top universities are establishing data-science labs and programs.
(There are many alternative names:
ML and statistics, computer science,
statistical learning, and data science,
among others.)

Financial applications of
machine learning
The area of finance has been relatively immune to the ML technology,

except for a few exceptions such as
high-frequency trading and credit
scoring for loans. Significant financial decisions are made with reference to formal decision models. The
first, and most famous, example is
the Markowitz portfolio model. This
model has been endlessly enhanced.
Institutional investors build structured asset allocation and asset-liability management models for a variety of reasons, including the desire to
improve performance and others,
such as satisfying governance procedures. Regulators may charge institutional investors with imprudence if
there is no formal analysis of their
portfolio selection decisions and large
losses occur. Due to the levels of
uncertainties and time-lags in strategic decisions, it can be difficult to
evaluate the quality of the recommendations, thus complicating the
search for ML breakthroughs. Recall
the need for measurements of correctness in supervised learning.
Still, there are a few applications.
One involves determining factors that
impact the performance of asset categories, especially the so-called alternative assets. Many institutional

FIG2 Factors underlying asset performance can be interpreted as ingredients for evaluating risks and returns (similar to ingredients in a cake recipe).

■

N o v e mber /D ecember 2017

IEEE POTENTIALS

investors have shifted capital to alternatives such as hedge funds, private
equity, credit funds, real assets, and
others. Evaluating the performance
of these assets is more complex than
traditional stocks and bonds due to
the presence of multiple risk factors.
Think of factor selection as the task
of choosing the underlying ingredients in a recipe such as in a cake
mixture (Fig. 2). The selection of factors and their weightings are identical to the feature selection process
in ML. The goal is to measure risks
more precisely than via traditional
portfolio approaches and to improve
investor performance through greater diversification. Many large institutional investors are turning to factor
investing approaches.
As an example, we employ the
following five core factors: world
equity, U.S. Treasury bonds, highyield bonds, inflation protection, and
currency protection. Next, we calculate the impact of these factors on the
performance of generic asset categories. The corresponding factor loadings are determined by reference to
a cross-validation approach in conjunction with a regularization term
(the Lasso) as a shrinkage penalty
on the size of the loadings. This approach is more robust than traditional regression models and is standard
in machine learning. Figure 3 shows
an illustrative case study.
A portfolio decision model takes
these factor loadings as input in its
search for determining an optimal
asset allocation or asset-liability solution. Risks are diversified across the
underlying factors, rather than solely
by means of the asset categories. The
goal of factor investing is to protect
the investor's capital in a comprehensive fashion, especially during market crashes.
A second application of ML in financial planning involves the identification of economic regimes. Here,
we partition a historical time period
into "regimes," each regime possessing
relatively homogenous patterns. You
can interpret the regimes as conditions of the market under various
environments, similar to the traffic
patterns during normal periods and

Table of Contents for the Digital Edition of Potentials - November/December 2017