ANN Approach to Speculate Stock Performance for Inter-Day Traders ECE 539 Chris Churas 5/05/2000OutlineProblem DescriptionData AcquisitionMethod OverviewNeural Network DesignANN InputsANN 5 OutputsANN 3 OutputsANN ConfigurationSlide 11Slide 12ResultsConclusionsANN Approach to Speculate Stock Performance for Inter-Day TradersECE 539Chris Churas5/05/2000Outline•Problem Description•Data Acquisition•Method Overview•Neural Network Design•ANN Inputs & Outputs•Performance and Results•ConclusionsProblem DescriptionGiven information pertaining to a particular stock. I want to be able to predict whether a stock will increase or decrease in price.The time interval of prediction will be between one and twenty minutes. The information will include stock price, current trading volume, P/E ratio and other factors.Data Acquisition•Gathered data from http://entrypoint.com–Downloading of data from entrypoint •Wrote Perl program getstocks.pl to download and parse stock data. This program also stored stock data in Mysql Database•The Perl programs getdata.pl and setdata.pl took data from Database and wrote to ascii file for input into Neural NetworkMethod Overview•Use of 2 Layer Artificial Neural Network•Input various stock attributes into ANN•Output binary value that denotes whether stock will increase or decrease in futureNeural Network DesignHidden Layer 8 NeuronsInput Layer 11 NeuronsOutput Layer 5 NeuronsUsing Sigmoid Activation FunctionDirection of ANNANN Inputs•Earnings Per Share•P/E Ratio•Avg. Daily Volume•Common Shares Out•52 Week High•52 Week Low•Day High•Day Low•Previous Day Close•Current Volume•Last PriceANN 5 Outputs•The ANN output is 5 binary bits. Which translate to the following:1 0 0 0 0 = Decrease in stock price > threshold0 1 0 0 0 = Decrease in stock price <= threshold0 0 1 0 0 = No change in stock price0 0 0 1 0 = Increase in stock price <= threshold0 0 0 0 1 = Increase in stock price > thresholdNote: Threshold is set by user in setdata.plANN 3 Outputs•The ANN output is 3 binary bits. Which translate to the following:1 0 0 = Decrease in stock price0 1 0 = No change in stock price0 0 1 = Increase in stock priceANN Configuration•ANN Parameters Used–Learning Rate: 0.5–Momentum: 0.9–Epoch: 500–Threshold: 0.15% of stock price•Data Statistics Per Stock–Number of Training Samples: 1200-1500–Number of Testing Samples: 300Classification Rates 11 Inputs 5 Outputs56%26%40%23%34%26%31%29%0%10%20%30%40%50%60%WDC WDC RAND AMD AMD RAND AOL AOL RAND TDFX TDFX RANDStocks in blue and Random Prediction in magentaPercentage CorrectClassification Rates 11 Inputs 3 Outputs56%37%41%34%36%34%34%33%0%10%20%30%40%50%60%WDC WDC RAND AMD AMD RAND AOL AOL RAND TDFX TDFX RANDStocks in blue and Random Prediction in magentaPercentage CorrectResults•5 Output ANN performance marginal– Only predictions on WDC stock had satisfactory results•3 Output ANN did not fair much better–Once again WDC was only stock to be predicted with some accuracyConclusions•Perhaps with improved data set that was not missing so much data all stocks could have results similar to WDC stock•If the performance could be improved to that of WDC stock. Then ANN could be implemented as part of a stock ticker for Inter-Day Stock
View Full Document
Unlocking...