Using Data Mining Technologies to find

Currency Trading Rules

A. G. MalliarisM. E. Malliaris

Loyola University Chicago

Multinational Finance Society, Rome, Italy, June 26-29, 2011

MOTIVATION

• If it is possible to isolate some assets that have consistent patterns of co-movement, then this knowledge can be used in two ways:

to make money if one has moved and the other has not moved yet, or,

to build a more diverse portfolio by including assets that move in opposite directions

FOCUS

• Are there certain currency markets that move up or down together sufficiently often for us to form a conclusion about their inter-relationships?

• Data from eight major currencies, over ten years, was studied to see if there are movement rules among these currencies that might be stable over time.

DATA MINING

• Data Mining is the search for meaningful patterns in large data sets

• Meaningful patterns are easily understood, valid on new or test data with some degree of certainty, and potentially useful

• Models are produced using a set of values of indicators at a particular time

• The goal is to produce a model that generalizes well on future observations

ASSOCIATION ANALYSIS

• Association analysis is a popular data mining method that originated with the study of market baskets to see which items people purchased at the same time.

• Association analysis generates a set of rules of the form IF A THEN B

ASSOCIATION ANALYSIS

• The set of rules that is generated also have values of support and confidence

• Support: percent of times that some combination of inputs (also called antecedents) occurs in the data set.

• Confidence: when the antecedent combination does occur, reflects the percent of time that the output, or consequent, is also true.

GRI

• Generalized Rule Induction is an association analysis methodology that was introduced in 1992 by Smyth and Goodman

• GRI is an effective, parsimonious method for detecting relationships in a large set of variables

• Decision trees divide up a large collection of records into successively smaller sets of records by applying a sequence of simple decision rules.

• A good decision tree model consists of a set of rules that results in homogeneous groups; that is, it separates records into groups where a single class predominates for each group

• The final result of these splits is often represented graphically in a tree structure.

DECISION TREES

All Data

Begin

• All Decision Trees begin with a root node.• They employ a strategy that grows the tree by

making a series of locally optimum decisions about which attribute to use for partitioning the data.

• The goal of the algorithm is to partition the records into successively purer subsets based on the values of the target field.

GROWING A TREE

CLASSIFICATION & REGRESSION TREES (C&RT)

• Unlike many methods from statistics, C&RT did not exist before machine learning methods were available.

• C&RT is a decision tree methodology that generates only binary splits at each stage

• All decisions are based on the value of a single target variable

DATA

• January 2000 through July 2009 • Downloaded from Bloomberg.• These prices were split into two disjoint sets for

training and validation. • Data from January 1 2000 to June 30 2008 was

used as the training set (2215 rows)• July 1 2008 to July 21 2009, used as the validation

set (276 rows). • To study the simultaneous market movements, all

data was transformed into “Up” or “Down”

THE EIGHT CURRENCIES

Original data was daily cash closing prices for the price of 1 US Dollar in the foreign currency that day

Australian Dollar, Japanese Yen British Pound, Euro, Swiss Franc Canadian Dollar, Mexican Peso, Brazilian Real

RELATIVE MOVEMENT

In order to view them all in a similar scale, the Mexican Peso has been multiplied by 10 and the Japanese Yen by 100 for the graph.

DATA MINING TOOL• Methodologies were run using the SPSS product

Clementine • There were two runs of each model for GRI

– First, the Australian dollar and Japanese Yen were inputs, with the Euro, the Swiss Franc and the British Pound as possible outputs.

– Second, the Australian dollar, Japanese Yen, the Euro, the Swiss Franc and the British Pound were inputs, with the Mexican Peso, the Brazilian Real and the Canadian dollar as outputs.

• One C&RT tree was created for each of the six GRI output targets

RESULTS

• Each model developed a large number of rules and paths.

• Here, we display one Up rule and one Down rule for each of the target markets.

• Rules selected were those that did well not only on the training set, but also on the validation set.

Training Set Validation Set

Target Rules With Best Results %Rows %Correct %Rows %CorrectEuro = Up Australia = Up and Japan = Up 28.94 78 17.39 85.42

Euro = Down Australia = Down and Japan = Down 32.42 81.45 21.84 85.71

Britain = Up Australia = Up and Japan = Up 28.94 72.07 17.39 77.08

Britain = Down Australia = Down and Japan = Down 32.42 76.71 21.84 76.19

Swiss = Up Australia = Up and Japan = Up 28.94 76.44 17.39 87.50

Swiss = Down Australia= Down and Japan = Down 32.42 80.22 21.84 80.70

Brazil = Up Australia = Up and Japan = Down 16.91 58.62 31.42 74.39

Brazil = DownAustralia = Down and Euro = Down and Swiss = Down and Britain = Down 27.8 61.45 23.75 80.64

Mexico = UpEuro = Up and Britain = Up and Swiss = Down 3.02 64.18 5.8 87.50

Mexico = DownAustralia = Down and Japan = Up and Britain = Down 11.53 63.03 18.77 83.67

Canada = Up Australia = Up 45.92 65.72 47.13 78.05

Canada = Down Australia = Down and Euro = Down 38.12 72.24 37.16 83.51

CONCLUSIONS

• Today, the daily volume of currency transactions in currency futures, forwards, swaps and options dominates all other types of trading volumes.

• Whether currencies move together or independently is a matter of importance for investors wishing to spread the impact of their portfolio decisions.

• The results of this study suggest that there is reason to believe that co-movement among some specific markets exists over relatively long periods of time.