Adaptive Trading System (ATS)

Goal of the system:

The objective of the system is to provide a recommendation on the purchase, sale, or retention of generic securities on a single day based on a prediction as whether the next day's closing value of a specific market index would be respectively higher than, lower than, or equal to that of the present day.

Trading model and assumptions:

The system recommends a trading decision based on a single market index, specifically Dow Jones Industrial Average (DJI). The trading interval is a single day. In specific, in the case of a 'purchase' recommendation, a security is purchased on the day of the recommendation and sold on the following day; in the same way, in the case of a 'sale' recommendation, a security is sold on the day of the recommendation and purchased on the following day; in the case of a 'retain' recommendation, no trading is performed. The trading model does not include short sales. The system uses the historical closing value of DJI in determining its trading decision.

Trading techniques:

The system performs data mining to classify the next day's closing value of DJI as higher than, lower than, or equal to that of the present day by extracting a pattern from the historical DJI closing value data using a set of technical analysis techniques as data attributes. In specific, simple moving average (SMA), exponential moving average (EMA), volume simple moving average (VSMA), double crossover, moving average convergence-divergence (MACD), and relative strength index (RSI) account for the trending and mean-reversion techniques used as data attributes.

Dataset sample:

@relation dji

@attribute SMA {rising,falling,flat}

@attribute EMA {rising,falling,flat}

@attribute VSMA {rising,falling,flat}

@attribute DoubleCrossOver {GoldenCross,DeadCross,NoCrossOver}

@attribute MACD {positive,negative,centerline}

@attribute RSI {overbought,oversold,neither}

@attribute classIndex {purchase,sale,retain}

@data

falling,falling,rising,NoCrossOver,negative,oversold,retain

falling,rising,rising,GoldenCross,negative,oversold,purchase

rising,rising,falling,NoCrossOver,negative,neither,retain

rising,rising,falling,DeadCross,positive,overbought,sale

Evaluation metric:

Cross validation between the training and test instances is performed on a complete data set using an untrained classifier. The percentage of correctly classified instances and kappa statistic are used as two evaluation metrics.

kappa = (P(A) - P(E)) / (1 - P(E)), where P(A) is the percentage agreement between the classifier and ground truth, and P(E) is chance agreement. k=1 indicates perfect agreement, and k=0 indicates chance agreement.

Machine learning algorithm and platform:

The following classifiers are used as the system's machine learning algorithms:

Decision Tree classifiers,

(A) C4.5 decision tree

(B) NBTree, a decision tree with Naive Bayes classifiers at the leaves.

(C) RandomForest, a random forest of 10 decision trees, each constructed while considering 3 random features.

Bayes classifiers,

(D) BayesNet, a Bayesian Network classifier.

(E) NaiveBayes, a Naive Bayes classifier.

Rules classifiers,

(F) DecisionTable

(G) ZeroR

(H) Ridor, Ripple Down Rule Learner

Meta Classifier,

(I) AdaBoostM1

The system is developed in Java using Weka as an external machine learning library.

Data source:

Historical closing value record of DJI for the past 6 months, 1 year, 5 years, 10 years, 25 years, 50 years, 70 years, and 82 years are obatined from Yahoo Finance as the source data.

Data usage:

Example-wise randomization is performed on the source data set. Seventy percent of the randomized data set is used as the training set and the remaining thirty percent as the test set.

Error function:

Root mean squared error is used as the error function in evaluating each machine learning experiment.

Review of the system:

The initial implementation of the system resulted in classification performance comparable to chance prediction. As a result, a review was conducted to determine whether (i) a technical problem exists in the implementation of the system or (ii) the set of technical analysis techniques used as dataset attributes failed to provide a statistically significant characterization of dataset instances.

No technical problems were found in the implementation of the system. The generation of training and test datasets were traced and verified at each step of the code. Sanity check was performed to verify the correct configuration of the ML algorithms using datasets for which the calssification results could be verified by hand. Specifically, all ML algorithms scored 100% in correctly classifying instances of datasets in which all instances had their ClassIndex attribute set to one identical nominal value, for example "retain".

The attributes (SMA, EMA, Double Crossover, MACD, and RSI) previously used to define each instance were all derived from the closing DJI value. Volume Simple Moving Average (VSMA), derived from the numbers of shares traded on each day, is also added to the set of attributes used to define each instance. Nevertheless, the new set of attributes (SMA, EMA, VSMA, Double Crossover, MACD, and RSI) still did not uniquely characterize a statistically significant number of instances to improve the classification performance of the system beyond predicition by chance.

As an alternate approach, in the second implementation of the system, a relationship between the attribute set and the ClassIndex was introduced to the dataset. Previsouly, the ClassIndex of each instance was set to "purchase", "sale", or "retain" exclusively based on whether the next day's closing value of DJI index was higher than, lower than, or the same as that of the present day. However, in order to introduce a relationship between the attribute set and the ClassIndex, the ClassIndex value of each instance was set not only based on the next day's closing value of DJI index compared to that of the present day, but, at the same time, based on the value of one or more technical analysis attributes (i.e. SMA, EMA, VSMA, DCO, MACD, and RSI) of that instance. For all ML algorithms used in this experiment, the highest classificaiton performance was achieved when the ClassIndex value was set according to the following conditions:

if ((Next day's closing value of DJI > Present day's closing value of DJI) &&

    (EMA = "rising" && VSMA = "rising") ||

    (Next day's closing value of DJI > Present day's closing value of DJI) &&

    (DCO = "GoldenCross" && VSMA = "rising")) ClassIndex = "purchase"

else if ((Next day's closing value of DJI < Present day's closing value of DJI) &&

    (EMA = "falling" && VSMA = "falling") ||

    (Next day's closing value of DJI < Present day's closing value of DJI) &&

    (DCO = "DeadCross" && VSMA = "falling") ClassIndex = "sale"

else ClassIndex = "retain"