Explorations of Machine Learning in Finance

[pic 1]

Department of Computer Science

Explorations of Machine Learning

in Quantitative Finance

Research Review

Lionel Cazarre-Cazalot

Supervisor: Dave Cliff

1. May 2nd 2014

   ---

2. Table of Contents

1        Introduction        2

1.1.        Motivation        2

1.2.        Refining the scope        3

1.3.        Aims and objectives        3

1.4.        Research methodology        4

2        Traditional financial valuation frameworks and their limitations        4

2.1        Quick review of traditional financial valuation frameworks        4

2.2        Limitations of the MPT and CAPM        5

2.3        General limitations of traditional models        5

3        Application of Machine Learning to Finance        6

3.1        Definition of Machine Learning        6

3.2        Taxonomy of ML algorithms        6

3.3        Survey of Machine Learning techniques applied to finance        7

3.4        Advantages of ML        8

3.5        Disadvantages of ML        9

3.6        Pitfalls of ML applied to stock market modelling        9

4        Evolutionary algorithms        10

4.1        Various types of Evolutionary Algorithms        10

4.2        Choosing among Evolutionary algorithms        11

5        Genetic Programming        14

5.1        The origins of GP        14

5.2        The structure and representation of GP programs        14

5.3        How GP works        16

5.3.1        The GP process in three steps        16

5.3.2        Generating the initial population        16

5.3.3        Fitness        17

5.3.4        Genetic operations        17

5.3.5        Termination criterion        18

5.4        Limitations of GP        18

6        Genetic Programmming applied to stock selection        18

6.1        Refining our objective        18

6.1.1        Stock investing        18

6.1.2        Return targets: absolute versus relative        19

6.2        Choice of the investable market        19

6.3        Candidate terminals        20

6.4        Data preprocessing        21

6.5        Candidate functions        22

6.6        Fitness function        23

6.7        Avoiding overfitting        23

6.8        Implementation        24

7        Data collection, preparation and preliminary analysis        25

7.1        Data requirements        25

7.2        Data providers        25

7.3        WRDS, Compustat and CRSP        25

7.4        Methodology used to build the dataset        26

7.4.1        Extraction of complete list of companies available from Compustat North America        26

7.4.2        Crosscheck of monthly stock prices and holding period returns against CRSP        27

7.5        The issue of missing data        27

7.6        Univariate linear regression analysis        28

7.7        Final dataset        29

7.8        Distribution of results        29

7.9        Multivariate linear regression analysis        30

8        Software implementation        31

8.1        Requirements        31

8.1.1        Business requirements        31

8.1.2        User requirements        33

8.1.3        Software requirements        34

8.2        Program design        35

8.3        Testing        36

8.4        Comment a sample run        37

8.5        Performance        40

9        Experiments        41

9.1        Objectives        41

9.2        Results        41

10        Process validation        44

10.1        Methodology        44

10.2        Results        44

10.3        Analysis        45

10.4        Caveats        46

11        Further work        46

11.1        Re-evaluating the results based on the statements publication date        46

11.2        Running the process on quarterly statements / evaluating new holding periods        47

11.3        Targeting risk-adjusted returns        47

11.4        Exploring more features        47