2017
Pattern Recognition in Stock Data
Kathryn Dover
Harvey Mudd College


Abstract iii
Acknowledgments xi
1  Introduction 1
2  Background 3
2.1  Methods for Pattern Recognition in Stock Data . . . . . . . . 3
2.2  StockPatterns ........................... 7
2.3  ProposedChangeinApproach ................. 10
3  Approach: Geometric Definition of Patterns 11
3.1 TheStandardW.......................... 11
3.2 TheStandardM.......................... 12
3.3 TheStandardHeadandShoulder................ 13
4  Approach: Geometric Definition of Fuzzy Shapes 15
4.1 TheFuzzyW............................ 15
4.2 TheFuzzyM............................ 16
4.3 TheFuzzyHeadandShoulder ................. 17
5  Results: New Approach on Handling the Shapes 19
5.1  ChangeofBasis .......................... 19
5.2  FlippingaShape ......................... 20
5.3  SymmetricRepresentation.................... 21
5.4  FuzzySymmetry ......................... 21
5.5  Categorizing Shapes Using Slopes and Lengths . . . . . . . . 22
5.6  RoughPredictions......................... 23
6  Implementation: Creating an Algorithm to Find Patterns 27
6.1 GaussianProcess ......................... 27
6.2 FindingLocalExtrema ...................... 29
6.3 CreatingVectorsUsingEndPoints ............... 29
6.4 StoringInformationforthePrediction . . . . . . . . . . . . . 29
6.5 RunningtheAlgorithmontheData .............. 29
7  Results: Running the Algorithm on Real Data 33
7.1 Results ............................... 33
7.2 Predictions............................. 36
7.3 PotentialIssues .......................... 40
8 Conclusion 45
8.1 FutureWork............................ 45
8.2 ClosingThoughts......................... 48
Bibliography 49  

Abstract
Finding patterns in high dimensional data can be difficult because it cannot be easily visualized. Many different machine learning methods are able to fit this high dimensional data in order to predict and classify future data but there is typically a large expense on having the machine learn the fit for a certain part of the dataset. This thesis proposes a geometric way of defining different patterns in data that is invariant under size and rotation so it is not so dependent on the input data. Using a Gaussian Process, the pattern is found within stock market data and predictions are made from it.