Understanding Why Artificial Neural Networks Fail in Finance (2013) | ||
Stock
markets. Bull and bear. Trading and investing. Lehman
Brothers (the “Bank
of Evil”). Underneath
these evocative terms and more is a complex, chaotic, non-linear time
series of prices. Their
complexity is such that retail investors may experience feelings of
disorientation with respect to exotic terms like derivatives,
OTC trading, low
latency trading, and dark
pools that seemingly require practitioners to don black robes and
chant around a mysterious circle. It
is supposed to be straightforward. Select
any accounting and finance text and be treated to chapters on the time
value of money, intrinsic monetary value, cash flow analysis, balance
sheet analysis, income statements, and more to help objectively evaluate
how much an asset is worth now from its likeliness of having some value in
the future. If the investor
calculates the asset is worth buying now at X, then they should buy if it
is on sale for less-than-or-equal to X and pass otherwise. Further treatment on fundamental and technical trading, especially with respect to the Foreign Exchange, can be found
here (updated 2020). Taking it to the other side raises the question of what the
counter-party values the same asset and why.
Adding in third party sell-side analyst reports does the same for
yet another dimension. How
much is the asset worth and who among us is “correct” or has the asset
“mis-priced?” At
the end of the day, one could apply this technique to buying clothing, or
a car, or a horse and carriage, or anything else at the local department
store or bazaar. Indeed,
markets and derivatives and hidden transactions predate by several
thousands of years the terms “Adjustable
Rate Mortgages, Collateralized Debt Obligations, and Credit Default Swaps”
involved in the latest 2007-2008 economic pain. Only the scale is larger, the layers a little deeper, and the
jargon more heavily specialized. The
result is a long time series that may not be truly random (Lo
& Mackinlay, 1988). To
help in their analysis, investors have supporting tools such as channel
and trend lines, various indicators, and pattern recognition techniques
to summarize the time series. These
tools can transform a price time series from this:
to
this:
which
has support and resistance lines, with some uptrending bottom-to-bottom
and top-to-top channel lines that implies the time series – Dow Jones
Industrial Average Q1 2013 – is rising.
These are classic linear tools.
There are also a myriad of non-linear, adaptive tools.
Some of the more famous non-classical non-linear and adaptive tools
are neural networks. Artificial
neural networks, as applied here as machine learning tools, are in the
broad gray-area intersection between what the brain does and how
(neuroscience) and what a computational algorithm could do (computer
science). Originating as a model of low level neuron anatomy (McCulloch
& Pitts, 1943) and scaling up to a model of self-programming
pattern recognition (Rumelhart,
et al, 1986) artificial neural networks have become a flexible and
powerful tool that greatly extends statistical regression into the layered
depths of non-linear and non-parametric data sets.
Various studies over the past two and a half decades mix variants
of artificial neural networks with financial time series data sets.
Bahrammierzaee
(2010)
compare artificial neural networks on three separate financial services
tasks, including credit evaluation, portfolio management, and financial
prediction. The neural
networks outperformed traditional statistical regression models in all
tasks, especially where the data became highly non-linear. Rezaiedolatabadi,
et al (2013)
use a hybrid mixture of different artificial neural network techniques to
enhance market prediction on the Tehran Stock Market, outperforming
individual techniques. Zhang
& Wu (2009)
borrow a technique termed improved bacterial chemotaxis – a variant of a
stochastic, genetic evolutionary algorithm – to blend with an artificial
neural network to enhance its stability, efficiency, and generalization.
These
are very clever and advanced mixture models applied to a complex time
series task. It appears to
make perfect sense – if the market produces stochastic volatility, the
neural network prediction model counters with wavelets and Boltzmann
Machines. It could very well
be that few if anyone on this planet can do better in each of these
particular micro-scale techniques. But
it must be remembered that artificial neural networks originated as core
models of what the brain does and how.
The intersection with computational algorithms and subsequent
merging into machine learning only arose and lasts with the need for an
objective and testable expression tool.
Artificial neural networks can and do perform complex pattern
recognition. But that only
scratches the surface of what they represent.
Applying neural networks to financial prediction solely because
their advanced non-linear pattern recognition are among the few that fit
chaotic price time series misses the main point.
This would be like an amateur chef stewing a truffle
– it turns a rare, expensive, and complex ingredient into common canned
goods. Truffles happen to be
mushrooms, but a mushroom is not a truffle.
A neural network can do pattern recognition, but pattern
recognition is not a brain model. This
is especially poignant given that the financial time series dataset is not
a set of data points. Every
tick on a stock price is in reality a buyer agreeing with a seller on the
price. The prices are an
aggregate transaction log, a history of decisions and communications among
the active traders and investors in the marketplace.
These are not simply arbitrary test points on a physical energy
gradient or plane. This is
not the XOR
problem. This is not simply
about covariance matrices and convexity of returns.
These are the beginnings and ends of hopes and dreams for the
future. These spring from the
brain. That taking a
byproduct of a brain model turned neural network and applying it to the
symptoms of human behavior as tracked by financial time series works is a
testament to the dedication and hard work in both fields.
But there is room for something greater. What
if the core neural network of the brain can directly merge with the cause
of human behavior in a higher-level model?
If human environments in real life are more about interactive
marketplaces than sterile lab data points translated into computational
mathematic algorithms, then perhaps the neural networks of the brain can
merge with financial analysis of human decision making under uncertainty:
Neural networks paired with finance but with cognitive and
neuroscience in the driver’s seat rather than machine learning and
stochastic math. Focusing on
such topics as selective attention,
executive attention, and mirroring neurons, such a partnership would
be behavioral finance on steroids meets executive cognitive development. Therefore,
fitting artificial neural networks with financial dataset does makes
perfect sense – but very importantly, not simply for the non-linear
machine learning capabilities that are advertised.
|