See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220835450

Algorithmic Trading Using Intelligent Agents.

Conference Paper · January 2008

Source: DBLP

CITATIONS READS
3 995

2 authors, including:

Orlando Belo
University of Minho
195 PUBLICATIONS   394 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

RAIDB2K View project

Using Process Mining on the Identification of ETL Black Points View project

All content following this page was uploaded by Orlando Belo on 16 May 2014.

The user has requested enhancement of the downloaded file.

 Algorithmic Trading Using Intelligent Agents
Rui Pedro Barbosa and Orlando Belo
Department of Informatics, University of Minho, Braga, Portugal
{rui.barbosa, obelo}@di.uminho.pt

professionals in the investment industry. That is due to the

Abstract - Trading in financial markets is undergoing a
fact that these studies usually measure a model’s
radical transformation, one in which algorithmic methods
performance based on its accuracy (for classification) or the
are becoming increasingly more important. The
mean squared error (for regression). From a trader’s point of
development of intelligent agents that can act as
view, there is an obvious problem with this approach: higher
autonomous traders seems like a logical step forward in this
accuracy does not necessarily translate into higher profit. A
“algorithms arms race”. In this paper we describe an
single losing trade can wipe out the profit of several
infrastructure for implementing hybrid intelligent agents
accurately predicted trades. A low mean squared error is
with the ability to trade financial instruments without
also not a guarantee that a model can produce profitable
requiring human supervision. We used this infrastructure,
predictions [6]. Some studies try to overcome this problem
which relies heavily on artificial intelligence models and
by using model predictions on out-of-sample data to
problem solving methodologies, to implement two currency
simulate trades. This is obviously a much better way to
trading agents. So far these agents have been able to profit
analyze the profit potential of a model, but it is still not a
from inefficiencies in the currency market, which lead us to
perfect solution. Trading simulations do not take into
believe our infrastructure can be of practical interest to the
consideration the costs of trading, such as commissions and
investment industry.
interest payments. They also do not take into account the
Keywords: Algorithmic Trading, Hybrid Agents. problems that can occur while trading live, such as slippage
and partial fills. The impact of these problems on the overall
profitability of a trading strategy is not negligible.
1 Introduction All things considered, the performance gauges that most
matter to the trading community are probably the profit and
Algorithmic trading is a form of automated trading that
maximum drawdown. The importance given to the profit is
consists in the use of computer programs for placing trading
rather obvious. The maximum drawdown, although a less
orders, with the computer algorithm deciding on the
known concept, is equally important. It is defined as a
different characteristics of the order, such as the size and the
trader’s worst period of "peak to valley" performance in the
price. The adoption of algorithmic trading by participants in
past. In other words, it is the maximum loss experienced by
different financial markets is growing at an extremely fast
a trader in the past, until a new high in the profit was
pace. This growth is particularly noticeable in the currency
achieved. So the drawdown can be looked at as a measure of
market, also known as Foreign Exchange Market or Forex
risk: the higher the maximum drawdown, the riskier the
Market, where the use of this novel form of trading is
trading strategy will be considered.
expected to grow from 7% by the end of 2006 to 25% by
2010 [1]. The development of intelligent agents that are able With the needs of the trading community in mind, in this
to act as autonomous traders is an obvious step forward in paper we will describe an infrastructure for implementing
this move away from traditional methods. With this in mind, trading agents whose main goal is to maximize the profit
in this paper we will describe a mechanism to implement and to minimize the drawdown while trading live. Unlike
autonomous Forex trading agents using an infrastructure most studies in this field, which describe tools that can be
entirely based in artificial intelligence methodologies. The used to aid the traders, we will be looking at a solution that
idea of using artificial intelligence models in trading is not can actually replace the traders. Therefore, each
new, as there are already numerous studies in this field. implemented agent is expected to be able to operate
Neural networks, for example, have been shown to be able autonomously, placing trades and handling money
to model financial time series better than traditional management without requiring human intervention. The
mathematical methods [2][3]. The use of hybrid intelligent infrastructure is loosely based in the decision process of a
systems for financial prediction has also been extensively traditional trader: it enables the agents to intuitively
researched, with several studies showing that, in general, recognize patterns in financial time series, to remember
hybrid systems can outperform non-hybrid systems [4][5]. previous trades and use that empirical knowledge to decide
when and how much to invest, and to incorporate
Even though most studies seem to show that artificial
knowledge from trading books and trading experts into the
intelligence models can produce reasonably accurate
trading decisions.
financial predictions, that in itself will not impress most
2 Forex Market
The Forex Market is the largest financial market in the
world. In this market currencies are traded against each
other, and each pair of currencies is a product that can be
traded. For instance, USD/JPY is the price of the United
States Dollar expressed in Japanese Yen. At the time of
writing of this paper the USD/JPY price is 104.32, meaning
we need 104.32 JPY to buy 1 USD. Trading this pair in the
Forex Market is pretty straightforward: if a trader believes
the USD will become more valuable compared to the JPY
he buys USD/JPY lots (goes long), and if he thinks the JPY
will become more valuable compared to the USD he sells
USD/JPY lots (goes short). The profit/loss of each trade can
be expressed in pips. A pip is the smallest change in the
price of a currency pair. For the USD/JPY pair a pip
corresponds to a price movement of 0.01. The actual value
of each pip depends on the amount invested. For example, if
we buy/sell 100,000 USD/JPY each pip is worth 1,000 JPY
(100,000 times 0.01), or 9.59 USD (1,000 divided by
104.32). Even though the Forex Market is well known for
its unpredictability, its high liquidity and volatility make it Figure 1. Infrastructure for implementing trading agents.
an interesting target for autonomous agents.
3.1 Intuition Module
3 Infrastructure
In order to trade, an agent must be able to decide when it
An autonomous trading agent will need to make several should buy or short sell a financial instrument. To do this, it
decisions before it can place a trade. For example, it will needs a way to intuitively predict if the price of the financial
need to decide if it should buy or sell a financial instrument, instrument will be going up or down in the future. In a very
how much it should buy or sell, and when it should close an simplistic way, we can consider that the intuitive capacity of
open trade. We designed an infrastructure that can be used a trader is somewhat similar to a complex pattern
to implement trading agents with the ability to make those recognition process. With this in mind, our agents’ intuitive
decisions. This infrastructure is represented in Figure 1. It predictions will be performed by classification and
defines two percepts (price changes over a period of time regression models capable of finding hidden patterns in
and result of previous trades) and a single action (placement financial data.
of new trades), and is composed of three interconnected
modules: Instead of using a single classification or regression model
to implement the agents’ Intuition Module, we use an
Intuition Module – this module is responsible for Ensemble Model, which should improve the prediction
predicting if the price of a currency pair will go up results [7]. This Ensemble is basically a weighted voting
or down. This prediction is done by an Ensemble system composed of several classification and regression
Model, which consists of several classification and models, where the weight of each vote is based on the
regression models that try to find hidden patterns in profitability of each model. The models do not try to predict
price data. the price in the future, they simply try to predict what will
A Posteriori Knowledge Module – this module happen to the price in the future. The prediction of each
uses information from previous trades to suggest model thus corresponds to one of two classes: “the price will
when and how much to invest in each trade. This go up” or “the price will go down”. That is straightforward
suggestion is done by a Case-Based Reasoning for classification models, but for regression models we need
System. Each case in this system corresponds to a to convert the price prediction into one of the classes. That
trade executed by the agent and its final result is very easy to accomplish: if the predicted price is higher
(profit or loss in pips). than the current price then the predicted class is “the price
A Priori Knowledge Module – this module is will go up”, otherwise if it is lower than the current price the
responsible for making the final trading decision, predicted class is “the price will go down”.
using the prediction from the Intuition Module and To guarantee that the agents can keep learning as time goes
the suggestion from the A Posteriori Knowledge by, the models in the Ensemble will need to be retrained
Module. This decision is done by a Rule-Based with new data as it becomes available. However, we only
Expert System, which contains several rules want the retrained models to become a part of the Ensemble
regarding when to invest and when to close a trade. if they can perform at least as well as the original models.
To accomplish this, before each prediction the Intuition The weight of the vote of each of the individual
Module splits all the available instances into two datasets: models in the Ensemble is continuously updated,
the test set, consisting of the most recent N instances, and allowing the Ensemble to adapt to the market
the training set, consisting of all the instances left. The test conditions. As a model becomes more profitable,
set is first used to simulate trades with the predictions of its vote becomes more important.
each model in the Ensemble (for each test instance, if the The Ensemble Model can combine the qualities of
model predicts “the price will go up” then a long trade is the best models at predicting when the price will go
simulated, otherwise if it predicts “the price will go down” a up and the best models at predicting when the price
short sell if simulated). The results of these simulations are will go down. That is accomplished by using the
utilized to calculate each model’s overall profit factor: models’ long profit factor and short profit factor as
their votes’ weight.
(1) An Ensemble Model makes the trading strategy
more resilient to changes in market dynamics. If a
Next, each model in the Ensemble is retrained with the
model in the Ensemble becomes unprofitable, its
training set, and the same simulations and calculations are
vote continuously loses weight, up to a point where
performed with the test instances. The Intuition Module now
its predictions are completely ignored. Since this
has a way to select the models that will be used to make new
model will be trained with more data, it is likely
predictions: for each model in the Ensemble, if the profit
that it will end up being replaced with a more
factor of the retrained model is at least as high as the profit
profitable retrained version of itself.
factor of the original model, then the retrained model
replaces the original in the Ensemble. Otherwise, the Our algorithm optimizes profitability instead of
original is kept and the retrained model is discarded. accuracy. Obviously the learning algorithms used
to retrain the models still optimize their accuracy,
Once the updating of the models has been completed, the but the decision to actually make the retrained
Intuition Module needs to decide what will be the weight of models a part of the Ensemble Model is based
each model’s vote when calculating the Ensemble entirely on their profitability.
prediction. To do so, it uses the trading simulation results
obtained previously with the test data, and calculates each Retraining the models before each prediction is the
model’s long profit factor and short profit factor: key to the trading agents’ autonomy. The agents
can keep learning even while trading, because new
(2) unseen data eventually becomes a part of the
training set.
(3) However, this algorithm has one obvious problem. It uses
the simulated profitability displayed with the test data to
The long profit factor measures how profitable a model was decide if an original model should be replaced with a
when predicting price increases in the test period, while the version of itself trained with more data. This means it
short profit factor measures how profitable it was when selects models according to their test predictions, which
predicting price drops. One of these values will be the might lead to selecting models that overfit the test data. But
weight of the model’s vote when performing a new this is possibly not a very serious flaw, because models that
prediction: the long profit factor if it predicts “the price will overfit the data and are unprofitable will eventually be
go up”, or the short profit factor if it predicts “the price will replaced with retrained versions of themselves (that may or
go down”. If the weight is a negative number the Intuition may not overfit a different set of test data).
Module replaces it with zero, which effectively means the
model’s prediction will be ignored.
3.2 A Posteriori Knowledge Module
After all individual models have made their predictions, the
Ensemble prediction is calculated by adding the weights of After deciding if it should buy or sell a financial instrument,
the votes of all the models that predicted “the price will go the agent will need to make another equally important
up”, and then subtracting the weights of the votes of all the decision: how much should it buy or sell. In order to
models that predicted “the price will go down”. If the accomplish this, it will need a way to predict how profitable
Ensemble prediction is greater than zero then the Intuition a trade will be. Using this prediction, it can optimize the
Module’s final class prediction is “the price will go up”, profit by investing more when it expects a trade to be very
otherwise if it is lower than zero the final prediction is “the profitable, and investing less when it expects it to be less
price will go down”. profitable. One way to implement this form of money
management would be to double the investment when a
The purpose of the described algorithm is to allow an agent trade is expected to be very profitable, use a normal
to take advantage of the different characteristics of the investment amount if a trade is expected to have moderate
models in its Ensemble, and to make sure it can keep profit, and skipping trades that are expected to be
learning over time. Concretely, there are several reasons to unprofitable.
use this algorithm:
In order to determine the expected profitability of a trade, The primary example of such a report is the United States
we will be looking at the individual predictions of the Nonfarm Payrolls, or NFP, released on the first Friday of
models in the Ensemble. Intuitively, we might expect that every month.
the probability of a trade being successful will be higher if
Other more important rules are those where the settings for
all the models make the same prediction (all predict “the
take profit orders and stop loss orders are defined. These are
price will go up” or all predict “the price will go down”),
necessary so that the agents know when to exit a trade. A
compared to a trade where the models’ predictions are
take profit order is used to close a trade when it reaches a
mixed (some predict “the price will go up” and some predict
certain number of pips in profit, to guarantee that profit. A
“the price will go down”). Empirical evidence demonstrates
stop loss order is used to close a trade when it reaches a
that these expectations are well founded. Certain
certain number of pips of loss, to prevent the loss from
combinations of individual predictions really are more
widening.
profitable than others. Our agents’ money management
strategy is based on that empirical observation. Before each trade, the Rule-Based Expert System receives
the prediction from the Intuition Module and the suggested
We implemented the A Posteriori Knowledge Module using
investment amount from the A Posteriori Knowledge
a Case-Based Reasoning System where each case represents
Module. It then uses the rules defined by experienced traders
a trade previously executed by the agent. The following
to make the final decision regarding the trade direction,
information is contained in each case in the database: the
investment amount and exit conditions. The agents can be
predicted class, the trade result (profit or loss in pips) and
made completely autonomous by using a broker’s API to
the individual predictions from the models in the Ensemble.
send the final trade decisions directly into the market.
These cases are used to calculate the expected profitability
of a new trade. Before the trade is placed, the Intuition
Module makes the Ensemble prediction and sends the 4 Software Implementation
sequence of individual predictions from the models in the
We implemented the previously described infrastructure in a
Ensemble to the Case-Based Reasoning System. This system
software program that can be easily used to create trading
retrieves from its database all the cases with the same
agents. We named it iQuant, short for “intelligent
Ensemble class prediction and the same sequence of
quantitative analyst”. A screenshot is shown in Figure 2.
individual predictions. If it cannot retrieve a predefined
minimum number of cases, it removes the last prediction in
the sequence of individual predictions and retrieves the
cases again. This process is repeated until enough cases are
retrieved, at which point the overall profit factor of the
retrieved cases is calculated using Equation (1). That value
is the expected profitability of the trade, and is used to
decide how much to invest: if the expected profitability is at
least as high as a predefined value the agent doubles the
investment, if it is lower or equal to another predefined
value it skips the trade, otherwise it uses the regular
investment amount.
A new case is inserted in the Case-Based Reasoning System
database after a trade is executed and closed, hence
becoming available for the calculation of the expected
profitability of future trades.
Figure 2. The iQuant.
3.3 A Priori Knowledge Module
The iQuant handles the retraining and the predictions of the
No matter how “smart” the implemented agents are, there is models in the Intuition Module using the Weka data mining
still some trading knowledge they will not be able to pick up API [8]. Rules in the A Priori Knowledge Module are
from their empirical trading experiences. For this reason, the handled by the Drools engine [9].
module responsible for making the final trading decision
consists of a Rule-Based Expert System where important Building a trading agent using the iQuant software involves:
rules can be defined by trading experts. Defining the models in the Ensemble, and the
attributes used to train them.
Some of these rules can be quite simple. For example, we
may want the agents to skip trades in low liquidity days, Setting the parameters in the A Posteriori
such as those around Christmas or New Year’s Day. Or we Knowledge Module to determine when a trade
may want them to skip trades whenever major economic should be skipped or when the investment should
reports are about to be released, to avoid the characteristic be doubled.
chaotic price movements that happen right after the release. Setting the rules in the Rule-Based Expert System.
The iQuant uses the proprietary API of a broker to send Table 1. USD/JPY agent’s Ensemble.
orders directly into the market, but it can also be set to
Model Attributes Predict
perform predictions without sending orders. This allows us Instance-Based hour (nominal), day of week
to use it as an autonomous trader or as a tool to facilitate class
K* (nominal), MA(6), current class
trading. C4.5 hour (nominal), day of week
class
Decision Tree (nominal), MA(6), current class
RIPPER hour (nominal), day of week
5 Sample Trading Agents Rule Learner (nominal), current class
class
hour (nominal), day of week
We used the iQuant software to implement two agents, one Naïve Bayes class
(nominal), current return
to trade the USD/JPY and the other one to trade the Logistic hour (nominal), MA(6), current
EUR/USD currency pair. Each of these agents places a trade class
Decision Tree class
every 6 hours, from Sunday 18:00 GMT to Saturday 00:00 Instance-Based hour (nominal), day of week
return
GMT. K* (nominal), MA(6), current class
hour (numeric), day of week
Support Vector
(numeric), MA(10), MA(2), current return
5.1 Agents’ Intuition Machine
return
To prepare the Intuition Module of the currency trading
agents we had to select the classification and regression Table 2. EUR/USD agent’s Ensemble.
models to be inserted in their Ensemble Models. To Model Attributes Predict
accomplish this, we started by obtaining historical price data hour (nominal), day of week
Naïve Bayes class
from a market maker’s website [10]. We downloaded 4,100 (nominal), current return
candlesticks for each currency pair, comprising the period hour (nominal), day of week
Support Vector
(numeric), MA(2), RSI(11), return
from May 2003 to January 2007. A candlestick is a figure Machine
ROC(12)
that displays the high, low, open and close price of a hour (nominal), day of week
financial instrument over a specific period of time. We then CART
(nominal), 2nd lagged return, class
Decision Tree
used those candlesticks to calculate the price return over RSI(2), ROC(2), ROC(5)
each 6 hours period, given that this was one of the attributes Support Vector
hour (nominal), day of week
we inserted in the training instances. Of the available (nominal), MA(6), MA(4), MA(3), return
Machine
returns, 4,000 (corresponding to the period from May 2003 current return
hour (nominal), day of week
to December 2006) were used to train the models and the Least Median (nominal), 5th lagged return, 4th
remaining 100 (corresponding to the month of January Squared Linear lagged return, 3th lagged return, 2nd return
2007) were used to test the models. Regression lagged return, 1st lagged return,
current return
The models were trained with attributes such as the hour, Instance-Based hour (nominal), day of week
the day of the week, lagged returns and the current class or return
K* (nominal), current class
return. We also used other attributes regularly used in Gaussian Radial current return, hour (numeric), day
technical analysis by traditional traders, such as Moving Basis Function of week (nominal), MA(12), class
Averages (MA), the Relative Strength Index (RSI) and the Network ROC(4)
Rate of Change (ROC). Table 1 and Table 2 describe the
After completing the implementation of the Intuition
attributes used to train and test each model in the agents’
Module of both agents, we used the modules’ predictions to
Ensemble Models. While the classification models try to
simulate trades with out-of-sample data corresponding to the
predict the next class (“the price will go up in the next 6
period from February 2007 to April 2008. The accumulated
hours” or “the price will go down in the next 6 hours”), the
profit in pips over this period is displayed in Figure 3.
regression models try to predict the price return in the next 6
hours period, and that return is then converted to a class.
As previously mentioned, before each trade the Intuition
Module uses a fixed size dataset to test the models and
calculate their simulated profitability. We made the agents
use a test set consisting of the most recent 100 instances,
which could be considered insufficient. However, there are
two reasons for using such a small set of test data:
The shorter the test set, the faster the agents can
adapt to changes in price volatility, because the
weights of the models’ votes are based in their
simulated profitability using the test instances.
The shorter the test set, the faster the agents can
learn new patterns, because test instances will Figure 3. Intuition Module accumulated profit.
become training instances faster.
The total profit for the USD/JPY Intuition Module was 4,690 closed with the take profit order would only be closed when
pips after 1,364 trades, with a drawdown of 671 pips. The the 6 hours period ended and a new trade was opened.
EUR/USD profit was 3,804 pips after 1,403 trades, with a
Figure 5 shows the results of combining each agents’
drawdown of 1,243 pips. It is fairly obvious that these
Intuition Module and A Priori Knowledge Module to
trading strategies need improvement: the drawdown is too
simulate trades with the out-of-sample data. The USD/JPY
high and the profit curves are too erratic.
combination netted 4,772 pips of profit after 1,325 trades,
with a drawdown of 465 pips. The EUR/USD combination
5.2 Agents’ Empirical Knowledge obtained 4,416 pips after 1,363 trades, with a drawdown of
To prepare the agents’ A Posteriori Knowledge Module we 624 pips. These lower drawdowns are exactly what we
just needed to define a couple of user variables regarding required, but unfortunately there is also a big profit
money management. After a couple of trial and error tests, reduction if we compare these results with the ones obtained
we decided to use the following settings: with the combinations between the Intuition Module and the
A Posteriori Knowledge Module.
double the trade size if the overall profit factor of
the cases retrieved is greater or equal to 0.5;
skip the trade if the overall profit factor of the
cases retrieved is lower or equal to 0;
require a minimum of 3 retrieved cases before a
decision is taken.
Figure 4 shows the result of combining the Intuition Module
and the A Posteriori Knowledge Module of each agent to
simulate trades using the out-of-sample data. The USD/JPY
combination performed 996 trades, with a final profit of
8,897 pips and a drawdown of 1,103 pips. The EUR/USD
combination performed 1,031 trades, with a profit of 4,932
pips and a drawdown of 1,473 pips. Even though there was a
great increase in the profit and a decrease in the number of Figure 5. Intuition Module and A Priori Knowledge Module
trades, these strategies still need improvement because their accumulated profit.
drawdowns are too high.
5.4 Results
Through simulation, we have shown that each module makes
a different contribution to the profit and the drawdown. The
actual agents, implemented using the iQuant software,
consist of all the three modules working together. Figure 6
shows the simulated trading results for both agents. They
were able to add the A Posteriori Knowledge Module ability
to increase the profits to the A Priori Knowledge Module
ability to reduce the drawdown. The USD/JPY agent
obtained 8,308 pips of profit after 971 trades, with a
drawdown of 760 pips. The EUR/USD agent obtained 5,944
pips after 1,005 trades, with a drawdown of 798 pips.

Figure 4. Intuition Module and A Posteriori Knowledge

Module accumulated profit.

5.3 Agents’ Expert Knowledge

To prepare the agents’ A Priori Knowledge Module we just
had to define the rules in their Rule-Based Expert Systems.
We inserted rules to stop trading in the following low
liquidity days: Christmas, New Year’s and Good Friday. We
also added a take profit rule for each agent. The USD/JPY
agent was set to automatically close a trade if the price
moved 0.40% in the predicted direction, while the
EUR/USD was set to do the same if the price moved 0.20%.
Since we did not insert any other rules, a trade that was not Figure 6. Trading agents’ accumulated profit.
 Table 3 resumes the trading statistics for both agents and the divided by 1.5625, which is the current price of one EUR in
combinations of their modules. It is interesting to notice that USD), and each pip would be worth $6.40 (64,000 times
the Intuition Module of the USD/JPY agent can only predict 0.0001). Therefore, its profit would be $38,042, or 38.0%,
if the price will go up or down with 53.59% accuracy, while and its maximum drawdown would be $5,107.
the EUR/USD module can only do it with 52.82% accuracy.
As previously mentioned, simulated results can give us a
These percentages might seem too low, but we must keep in
general idea regarding an agent’s ability to be profitable
mind that these modules optimize profitability instead of
while trading live, but does not provide any guarantees. The
accuracy. Therefore, even though they are not very accurate,
only way to prove that an agent can actually be profitable is
the profit they obtain from accurately predicted trades is a
to let it create an extensive track record of live trading. In
lot higher than the losses suffered from incorrectly predicted
order to accomplish this, our iQuant agents have been
trades. The modules’ success rate, i.e., the percentage of
trading autonomously for several months using an Electronic
trades that are closed in profit, is equal to their accuracy
Communication Network. As expected, the agents’ live
because all the trades are closed at the end of the 6 hours
trading results are not as good as the simulated results, with
period, when a new trade is opened.
a decrease of around 25% in the total profit. This difference
Another interesting statistic in this table is the low number is due to commissions, slippage, partial fills and interest
of trades performed by the agents, when compared with the payments, amongst other things. However, these results are
number of trades that would be performed by their Intuition still very satisfactory. Even though it might be too soon to
Modules alone. That difference is due to the fact that the reach any conclusions, they seem to demonstrate the agents
agents’ A Posterior Knowledge Module and A Priori can profit from inefficiencies in the Forex Market.
Knowledge Module can make them skip trades that are
In order to reduce the trading risk, we could try to diversify
expected to be unprofitable. Also notice that both agents
the investment by using the iQuant software to implement
have a success rate that is considerably higher than their
agents with the ability to trade other financial instruments,
accuracy. That is due to the take profit rule in their A Priori
such as stocks or futures. We are currently looking into this
Knowledge Module. This rule allows the agents to be
multi-agent investment strategy, which we believe could be
profitable even if they make a wrong prediction.
of practical interest to the algorithmic trading community.
Table 3. Trading statistics.
Acc. Succ. Profit DD 6 References
Modules Trades
(%) (%) (pips) (pips)
Intuition 53.59 53.59 4,690 671 1,364
[1] Toby Cole. “Foreign Exchange Implications of
Algorithmic Trading”. Foreign Exchange Contact Group,
USD/JPY

Intuition +
56.33 56.33 8,897 1,103 996 May 2007.
APosteriori
Intuition +
53.43 55.02 4,772 465 1,325 [2] Christian Dunis and Mark Williams. “Modelling and
APriori
Trading the EUR/USD Exchange Rate: Do Neural Network
Agent 56.33 57.67 8,308 760 971
Intuition 52.82 52.82 3,804 1,243 1,403
Models Perform Better?”. Derivatives Use, Trading &
Regulation, No. 8/3, pp. 211–239, 2002.
EUR/USD

Intuition +
53.64 53.64 4,932 1,473 1,031
APosteriori [3] Joarder Kamruzzamana and Ruhul Sarker. “Comparing
Intuition + ANN Based Models with ARIMA for Prediction of Forex
53.19 57.96 4,416 624 1,363
APriori Rates”. ASOR Bulletin, Vol. 22, No. 2, pp. 2–11, 2003.
Agent 54.03 59.00 5,944 798 1,005
[4] Ajith Abraham. “Analysis of Hybrid Soft and Hard
To make the analysis of the results more intuitive, it might Computing Techniques for Forex Monitoring Systems”.
be interesting to see how the agents’ performance would Proceedings of the 2002 IEEE International Conference on
translate into actual money won or lost. Forex investments Fuzzy Systems, pp. 1616–1622, 2002.
are usually leveraged (which means they are done with [5] Lean Yu, Shouyang Wang and Kin Lai. “Designing a
borrowed funds), so the total profit obtained by an agent will Hybrid AI System as a Forex Trading Decision Support
always depend on the leverage it uses. Let us assume we had Tool”. Proceedings of the 17th IEEE International
a starting capital of $100,000, and using the rules in the A Conference on Tools with Artificial Intelligence, pp. 89–93,
Priori Knowledge Module we instructed our USD/JPY agent 2005.
to use a maximum 2:1 initial leverage, by setting its trade
[6] Kevin Swingler. “Financial Prediction, Some Pointers,
size to 100,000 USD/JPY. In this scenario, the agent only
Pitfalls, and Common Errors”. Neural Computing &
uses the maximum leverage when it doubles the trade size
Applications, Vol. 4, No. 4, pp. 192–197, 1996.
for trades with high expected profitability. For a USD/JPY
price of 104.32, the pip value for a 100,000 USD/JPY trade [7] Pádraig Cunningham. “Ensemble Techniques”.
is $9.59. Since our agent obtained a total profit of 8,308 Technical Report UCD-CSI-2007-5, April 2007.
pips, its profit in dollars after 15 months of trading would be [8] Weka API, http://www.cs.waikato.ac.nz/ml/weka/.
$79,674, or 79.7%, and its maximum drawdown would be [9] JBoss Drools, http://www.jboss.org/drools/.
$7,288. If the same scenario is applied to the EUR/USD
agent, it would use a 64,000 EUR/USD trade size (100,000 [10] OANDA Corporation, http://fxlabs.oanda.com.

View publication stats