Portfolio selection with classification trees - Semantic Scholar

2 downloads 0 Views 244KB Size Report
Mar 9, 2006 - (2000) use CART to partition assets into outperforming and .... Feb99 Aug99 Feb00 Aug00 Feb01 Aug01 Feb02 Aug02 Feb03 Aug03 Feb04 ...
Portfolio selection with classification trees M. Gilli

E. K¨ellezi

I. Roko

University of Geneva

Mirabaud & Cie

University of Geneva

CCFEA, University of Essex March 9, 2006

1

Motivation Classical approach balances risk and return, static, relies on past information, unable to capture variations in investment opportunities How to overcome these inconveniences? • Model future of the time-varying behavior of mean returns, variances and covariances • Factor models ◦ statistical (black box) ◦ macroeconomic ◦ fundamental Gilli, K¨ellezi and Roko

CART – 2

Outline • Factor models for portfolio selection • Decision trees • Application S&P technology (1999–2005) • Conclusions Gilli, K¨ellezi and Roko

CART – 3

Fundamental factor models Returns on assets ri explained by characteristics ( z1 , z2 , . . . ) of individual assets: • • • • •

Balance sheet average Liquidity Market capitalization Price/earning ratio Earnings growth

How to model r

i,

t+1

= f (z

1

t

,z

2

t

, . . .)

Linear models will generally not work. Gilli, K¨ellezi and Roko

CART – 4

2

Modelling approaches If rit are discrete (discretized) we can use: • • • •

Ordered logistic or Probit regression Artificial Neural Networks (ANN) Genetic algorithms Recursive partitioning algorithms CART ( C lassification A nd R egression T rees) Breiman et al. (1984)

Able to handle highly nonlinear models. Gilli, K¨ellezi and Roko

CART – 5

Advantages of decision trees • Provide interpretable rules and logic statements • No assumptions about the statistical distributions of variables • Provides a clear indication which variables are most important for classification • Robust with noisy data – outliers, errors, etc. Gilli, K¨ellezi and Roko

Decision trees • • • • •

CART – 6

(applications)

Medicine: identification of factors influencing myocardial infarction, drug testing, etc. Artificial intelligence (machine learning) Internet engine Database screening Management and finance: • •

Frydman et al. (1985): default risk personal loans, credit rating, asset allocation, stock screening

Gilli, K¨ellezi and Roko

CART – 7

Decision trees for portfolio selection • Kao and Shumaker (1999) use CART to explain relationships between macroeconomic variables and performance of timing strategies based on market, size, and style. • Sorensen et al. (2000) use CART to partition assets into outperforming and underperforming assets. Portfolio composed by uniformly weighted outperforming assets. • Albanis and Batchelor (2000) compare different techniques to distinguish outperforming and underperforming assets and demonstrate the efficency of CART. Gilli, K¨ellezi and Roko

CART – 8

3

Decision trees for portfolio selection

(contd.)

We build on Sorensen et al. (2000) • Use a moving window to build successive classification trees • Construct portfolios and analyse their performance including non-convex transaction costs Gilli, K¨ellezi and Roko

CART – 9

Building a classification tree y

Recursive partitioning algorithm

variable by building a binary decision tree according to some splitting rule based on discrete or continuous predicting variables Data separated into: learning and testing samples

x1 x2

y

x1 x2

O

1

1

4

N

2

1

1

N

2

4

2

U

4

1

5

U

3

4

2

U

9

1

4

U

4

3

3

U

3

2

1

N

5

2

4

N

5

2

4

O

6

1

5

O

6

3

3

O

7

2

1

O

7

4

2

N

8

1

8

4

3

tx1 < N2.5 u 1

9 5 2 5u O 1 dependent explaining a discrete tx6U< 2.5 u tx22 < 3.5

tx3

Suggest Documents