A neural network is a mathematical modeling tool that has
the capacity to learn by example. This is an extraordinarily useful ability,
especially in financial modeling, where the there are usually countless
examples. Neural networks use a different technique from standard analysis that
is better suited to real world problems.
Networks first need to be trained by
being presented with hundreds or even thousands of facts, each fact consisting
of inputs and corresponding outputs. Through a unique feedback process, the
network learns how those inputs are related to the outputs, and develops a
general model that describes the relationship.
If training is successful, then it
will understand how to interpret new information. In all neural net training runs, some data is held back for
testing and additional validation. If
the net has not learned anything, and merely memorized its training data, then
it will score well when presented with the training data, but poorly when
presented with new fresh data. A
successfully trained net performs virtually the same on both sets.
Mathematical models are normally built by making a
priori assumptions about the functional form of the solution. These are
called parametric models, and are solved by regression methods to determine a
number of coefficients. This is sufficient if you know that the solution must be
a second order polynomial, or some other simple, well-known function. But in the
real world, relationships are not necessarily simple. Inputs and outputs could
even be related in a non-linear fashion. If you do not have to guess the
functional form of the answer, you have a big advantage.
Lets say you have three indicators that are predictive of
future price trends, A, B, and C. Empirically
you find that when A is up and B is down and C is up, that price tends to rise
for several weeks. One could set a
threshold or make a rule using A, B, and C, but by doing this you might miss
some signals. For instance, what if
each time A is up 50% more than B, the reading of C may need to be
down to get good signals. These
trade-offs and higher order combinations are very hard to ferret out of the
data. Neural networks are designed
to do this. If a network were
trained on many examples of A, B, and C, then only one rule would ever be
necessary, such as “buy when the neural net reading is greater than 0.95.
All possible trade-offs between A, B, and C that result in an up forecast
are being considered within the trained net.
The MM “T”
neural net uses eight indicators that are quantitative in nature.
They measure behavioral patterns such as trend persistency, unusual
volume activity, volume trends, price geometry, relative volatility, and
non-linear cycles. These indicators
are proprietary, and were built using the principles of Chaos Theory.
The goal of our behavioral studies is to find early telltale signs of
persistent positive investment behavior. The
most important manifestation of positive behavior is trend persistence, and the
associated expansion of investor interest as price increases. We use the Hurst exponent to measure trend persistency, and
volume accumulation to measure investor interest. Of course, the persistence of a trend is always threatened by
the tendency for investors to become overly exuberant. Every trend tends to go to an extreme, and evoke its own
reversal. Over indulgence and
exuberance are shown by extreme accelerating trend persistency, with associated
increases in volatility, and extreme turnover.
Our model is sensitive to these factors, and can anticipate reversals.
Trend lines and channels exist because investors make buy/sell decisions
as price contact these invisible price lines, thus creating and reinforcing them
through feedback. We map trend channels, because they provide feedback about
performance expectations. Within
the geometrical context of these channels, our studies measure the rates of
change and acceleration of trend persistency, volume accumulation, volatility,
price disparity with respect to expectations, and behavioral time cycles.
Chaos theory is the scientific theory that deals with
non-linear systems. These are
systems in which some of the outputs are fed back in to form the next set of
outputs. Auction markets, while not
being perfect non-linear systems, do demonstrate many of their characteristics.
Investors exhibit herd behavior on occasion.
By understanding these characteristics Marque Millennium has been able to
build better indicators than has previously been possible.
The “T” neural net was trained to find stocks that would outperform the S&P 500 index over a one-year holding period.
The fifteen factors listed below are used to train a series
of neural networks, which are then used to find the fair market value for a
Return on Equity
Growth in Sales
Adjusted Book Value
Cost of Capital
Industry Group (SIC)
Dividend Payout Ratio
Economic Value Added
Operating Income as % of Net Sales
Net Recurring EPS as % Sales
14 Cost of Goods Sold as % Net Sales
Net Recurring EPS as % Operating
The Marque Millennium equity-pricing model uses neural
networks to learn how a stock’s fundamental balance sheet data has been
translated to a market price in the past, within the context of its industry
group and economic sector. The models created for each industry group are then
used to estimate current fair market values for these stocks.
Equity screening has the
drawback of not considering all fundamentals and their relationships at once.
If a P/E is too high, a stock might be thrown out, while its sales, book,
and cash flow might be exceptional, and argue for its inclusion.
Our neural models will consider all factors simultaneously within the
context of the industry group.