Forex Trading Using Intermarket Analysis

map of a suCCessful neural network trading program. vantagepoint 
is an example of an analytiCal software program that uses multiple 
neural networks to analyze data and produCe market foreCasts.
F
igure
6.2.

---

t r a d e s e c r e t s
66
• The first network forecasts tomorrow’s high to help set
stops for entry and exit points.
• The second network forecasts tomorrow’s low to help set
stops for entry and exit points. 
• The third network forecasts a five-day moving average of
closes two days into the future to indicate the expected 
short-term trend direction within the next two days. 
• The fourth network forecasts a ten-day moving average of
closes four days into the future to indicate the expected 
medium-term trend direction within the next four days.
• The fifth network indicates whether the market is expect-
ed to change trend direction within the next two days, by 
making a top or a bottom.
The first four networks at the primary level of the network hierarchy 
make independent market forecasts of the high, low, short-term trend 
and medium-term trend. These predictions are then used as inputs 
into the fifth network, along with other intermarket data inputs, at the 
secondary level of the network hierarchy, to predict market turning 
points.
Once raw input data have been selected, it is preprocessed or mas-
saged using various algebraic and statistical methods of transforma-
tion, which help to facilitate “learning” by the neural network. That 
means it is converted into a form that the learning algorithm in the next 
layer can best exploit to get the most accurate forecasts in the shortest 
amount of time.
Hidden layer
The hidden layer is the learning algorithm used for internal process-
ing to store the “intelligence” gained during the learning process. 

---

67
ForeX trading using interMarket anaLysis
There are a number of learning algorithms. The network recodes the 
input data into a form that captures hidden patterns and relationships 
in the data, allowing the network to come to general conclusions from 
previously learned facts and apply them to new inputs. As this learn-
ing continues, the network creates an internal mapping of the input 
data, discerning the underlying causal relationships that exist within 
the data. This is what allows the network to make highly accurate 
market forecasts.
Many different learning algorithms can be used to train a neural net-
work in an attempt to minimize errors associated with the network’s 
forecasts. Some are slow while others are unstable.
Training a neural network is somewhat like human learning: repeti-
tion, repetition, repetition. The neural network learns from repeated 
exposures to the input data, and learned information is stored by the 
network in the form of a weight matrix. Changes in the weights occur 
as the network “learns.” Similar to the human learning process, neu-
ral networks learn behaviors or patterns by being exposed to repeated 
examples of them. Then the neural networks generalize through the 
learning process to related but previously unseen behaviors or pat-
terns. One popular network architecture for financial forecasting is 
known as a “feed-forward” network that trains through “back-propaga-
tion of error.” 
Although a neural network-based trading program can accommodate 
and analyze vast amounts of data, one thing a programmer must avoid 
is “over-training,” which is analogous to “curve-fitting” or “over-opti-
mization” in testing rule-based trading strategies. It takes considerable 
experimentation to determine the optimum number of neurons in the 
hidden layer and the number of hidden layers in a neural network.
If the hidden layer has too few neurons, it cannot map outputs from 
inputs correctly. If a network is presented with too many hidden layer 

---

t r a d e s e c r e t s
68
neurons, it memorizes the patterns in the training data without devel-
oping the ability to generalize to new data and discover the underly-
ing patterns and relationships. An over-trained network performs 
well on the training data but poorly on out-of-sample test data and 
subsequently during real-time trading—just like an over-optimized 
rule-based system.