portfolio selection: experimental comparison of universal and non-universal algorithms lorenzo...

43
PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information Processing, Spring 2011

Upload: ilene-howard

Post on 29-Dec-2015

218 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

PORTFOLIO SELECTION:EXPERIMENTAL COMPARISON OF UNIVERSAL AND

NON-UNIVERSAL ALGORITHMS

Lorenzo Coviello and Petros Mol

June 2, 2011

Universal Information Processing, Spring 2011

Page 2: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

2

Motivation

• Investing money in the stock market

• How to build a successful portfolio?

• Compare various strategies

Page 3: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

3

• Universal portfolio selection: provides guarantees on wealth growth rate

• Real market: invest in the most profitable way

• Compare performance of portfolio selection criteria on real data from the stock market

Introduction

Page 4: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

4

Rest of the talk • Introduction

• Portfolio selection: the model

• Methodology

• Two approaches• Reversal to the mean• Trend is your friend

• Simulations - Comparison

Page 5: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

5

The model – price relatives• Portfolio: m stocks

• Trading period: T trading days

• Xij: price relative of stock j at day i

• Xi often assumed i.i.d. (strong assumption)

Page 6: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

6

The model - wealth• Portfolio at day i

• The wealth gain in one day

• The overall wealth gain in T days

Page 7: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

7

The model - strategy• How to distribute the wealth among the stocks?

• Decision problem: choose a portfolio each day

Page 8: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

8

Rest of the talk • Introduction

• Portfolio selection: the model

• Methodology

• Two approaches• Reversal to the mean• Trend is your friend

• Comparison

Page 9: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

9

Methodology

• Data Collected from Yahoo! finance

• Adjusted close price used

• Period: 1996- 2010• 3778 trading days

• No priors on the stocks, no fundamentals

• No transaction costs

Page 10: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

10

Portfolio: List of Stocks

• Tech (11) : AMD, Apple, AT&T, Cisco, Dell, HP, IBM, Intel, Microsoft, Nokia, Oracle

• Finance (7): American Express, Bank of America, Barclay’s, Citigroup, JP Morgan, Morgan

Stanley, Wells Fargo

•Other (12) : Boeing , BP, Coca-Cola Company, Exxon, Ford, General Electric, J&J, McDonalds,

Pfizer, P&G, Wall Mart, Walt Disney

Page 11: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

11

Rest of the talk • Introduction

• Portfolio selection: the model

• Methodology

• Two approaches• Reversal to the mean• Trend is your friend

• Comparison

Page 12: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

12

Two main approaches

• Reversal to mean• Assume stock growth rates stable in the long run, and• Occasional larger returns followed by smaller rates• CRP, Semi-CRP, ANTICOR

• Trend is your friend• Portfolio based on recent stock performance• Histogram portfolio selection, kernel portfolio selection

Page 13: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

13

Buy and hold• Build portfolio once, let the wealth grow

• Uniform buy and hold (U-BAH)

• Performance guarantees for U-BAH

• Best BAH in hindsight: invest on the best stock

Page 14: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

14

Simulation

Page 15: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

15

Rest of the talk • Introduction

• Portfolio selection: the model

• Methodology

• Two approaches• Reversal to the mean• Trend is your friend

• Comparison

Page 16: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

16

Reverse to mean approach

Assumptions• Stock growth rates stable in the long run• Occasional larger returns followed by smaller rates, and vice

versa

Page 17: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

17

Constant rebalancing portfolio

• Rebalance portfolio every day according to pmf b

• Uniform CRP:

• Exponential gain if “reversal to the mean” market• Stock 1: constant value• Stock 2: doubles on odd days, halves on even days• Uniform CRP• Wealth grows of 1/8 every 2 days

• Best CRP in hindsight difficult to compute

Page 18: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

18

Semi-constant rebalanced portfolio

• Reference: Kalai (1998), Helmbold (1998), Kozat (2009)

• Portfolio rebalanced every arbitrary period

• Rebalancing period can be fixed

• Real market: reduced commissions

Page 19: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

19

Semi-constant rebalanced portfolio• Consider rebalancing every d days

• Uniform target distribution

• The wealth before rebalancing for the kth time

Page 20: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

20

Semi-CRP with deviation control

• Ref. Kozat (2009)

• Idea: avoid useless rebalancing

• Rebalance only if large distance between target portfolio b and current wealth distribution w

Page 21: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

21

Simulation (with fixed interval)

Page 22: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

22

Simulation (with distance threshold)

Page 23: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

23

ANTICOR algorithm

• Reference: Borodin, El-Yaniv, Gogan (2004)

• Aggressive “reversal to the mean”

• Transfer money from stock i to stock j if• Growth of stock i > growth of stock j over last window• Stock i in second last window and stock j in last window

positively correlated

Page 24: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

24

ANTICOR algorithm• Define

• Averages of columns of LXk

Page 25: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

25

ANTICOR algorithm• Cross correlation• stock i over the second last window• stock j over the last window

• Normalization

Page 26: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

26

ANTICOR algorithm• Transfer money from stock i to stock j if

• In an amount proportional to

Page 27: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

27

Simulation (with variable window length)

Page 28: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

28

Simulation (smaller window length)

Page 29: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

29

Simulation (zoom in)

Page 30: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

30

Simulation (zoom in)

Page 31: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

31

Simulation (zoom in)

Page 32: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

32

Rest of the talk • Introduction

• Portfolio selection: the model

• Methodology

• Two approaches• Reversal to the mean• Trend is your friend

• Comparison

Page 33: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

33

The trend is your friend

• Portfolio based on stock performance

• Prefer performing (trendy) stocks

• Use the market history to determine the current portfolio

Page 34: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

34

Histogram portfolio selection

• Ref: Gyorfi and Schafer (2003)

• Rectangular window of width w days

• Distribute the wealth uniformly among k best stocks

Page 35: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

35

Simulation (variant window)

Page 36: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

36

Simulation (variable #active stocks)

Page 37: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

37

Kernel portfolio selection• Higher weight to the recent past

• Window size of w days

• Window shape• Linear• Exponential

• Example: score of stock j at day i+1

Page 38: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

38

Kernel portfolio selection

• Each day the scores determine the portfolio

• Examples• Follow the best stock• Uniform distribution between k best stock• Proportional to score for best k stocks

Page 39: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

39

Simulation

Page 40: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

40

Summary of Cases

Reversal to the mean Trend is your friend- Constant Rebalancing (CRP)- Semi-CRP- ANTICOR

- Buy and Hold- Histogram- Kernel

Page 41: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

41

Comparing the winners (w/o Anticor)

Page 42: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

42

Conclusion

Put all your money in Anticor!

But choose the right window!!!

Page 43: PORTFOLIO SELECTION: EXPERIMENTAL COMPARISON OF UNIVERSAL AND NON-UNIVERSAL ALGORITHMS Lorenzo Coviello and Petros Mol June 2, 2011 Universal Information

THANKSLorenzo Coviello and Petros Mol