Portfolio-Optimization with Multi-Objective Evolutionary Algorithms in the case of Complex...
-
Upload
darren-york -
Category
Documents
-
view
217 -
download
1
Transcript of Portfolio-Optimization with Multi-Objective Evolutionary Algorithms in the case of Complex...
Portfolio-Optimization with Multi-Objective Evolutionary Algorithms in the case of
Complex Constraints
26.08.2005
Benedikt Scheckenbach
2
Outline
1. Basics of Portfolio Optimization2. Examples of Complex Constraints3. Existing MOEAs4. Idea of the thesis
3
Price of a share
• Price of a share can be regarded as a stochastic process.
• We define the return at a future date as
• Central Assumption: The returns are normally distributed
4
Definition of a Portfolio
• A portfolio is a bundle of shares.• Self-similarity property of Normal Distribution:
Returns of shares a normal distributed Return of portfolio is normally distributed.
• The money invested in each share is a portion (weight) between 0 and 100% of the portfolio price.
• The sum of all weights has to be 100%.
10%
20%
40%
5%
25%
Share 1
Share 2
Share 3
Share 4
Share 5
5
Why Portfolio Optimization?
• Diversification Portfolio might have lower variance than every single share.
• Individuality Each investor can adjust variance and mean to his needs.
• Simple Example: 2 Shares – Bivariate normal distribution of single returns.– Portfolio return is a convolution of single returns.
Correlation between two shares is important.
6
Diversification Effect in case of two shares
• Mean of Portfolio:
• Variance of Portfolio:
• Standard-Deviation:– 2 special cases:
7
Mean-Variance Portfolio Optimization
• „Classic“ optimization problem:
• Without further constraints there exists an analytical solution.
• In reality, further constraints have to be considered:– Additional requirements regarding the portfolio‘s weights.– Cardinality constraints.
E(x)
V(x)
Pareto Front
Minimum-Variance Portfolio
8
• In-house requirements:– Parts of the portfolio shall be invested in specific countries, sectors or
branches.– Each share is required to have a minimum weight to reduce transaction
costs (Buy-in threshold).
• Legal requirements:– German Investment Law §60 (1):
• The weight of each share has to be below 10%.• The sum of all weights above 5% may not exceed 40%.
Additional Requirements regarding the weigths
9
Cardinality Constraints
• Index-Tracking:– Financial products often have a share index as underlying.– Sometomes not all shares have an sufficient turnover volume.– To price the product one has to rebuild the index with only a few
shares. We need to find a portfolio that matches expected return and variance
of the index as close as possible with a maximum given number of shares.
or
10
Extended Optimization Problem
• Very large search space because of the combinatorial constraints. Application of MOEAs.
11
Existing MOEAs
• Focus on Cardinality Constraints, only buy-in thresholds as additional requirements regarding the weights.
• Phenotype: One Point in space
• Genotype: Mostly real-valued representation of weights.
• Non-dominated sorting according to NSGA-II.
• Critic:– Slow Convergence.– Algorithms don‘t incorporate special features of portfolio optimization.
• Critical Line Algorithm: Calculates the Pareto-Front for a given set of linear constraints.
12
Critical Line Algorithm (1)
• In the following: no cardinality constraints.• Input for Critical Line Algorithm: Concrete specification of basic
constraints as a system of linear inequalities.
• A and b specify linear constraints that fulfill basic constraints
Basic Problem Specification of basic Problem
13
Critical Line Algorithm (2)
• Example: Possible Matrix and RHS that fulfill German Investment Law:
14
Critical Line Algorithm (3)
• Output of Critical Line Algorithm: Weights of specific „Corner Portfolios“ that lie on Pareto Front for given constraints.
• All other portfolios of the Pareto-Front can be constructed as linear combinations of neighbored Corner-Portfolios.
15
Idea of the Diploma thesis
• Using Critical Line Algorithm as decoding function.• New geno- and phenotypes.
New non-dominated sorting, crossover, mutation.
Diploma thesis Present Algorithms
Genotype Set of linear constraints that fulfill basic constraints
Weights that fulfill basis constraints
Phenotype Complete Pareto-Front, i.e. best portfolios for given constraints
Single, suboptimal portfolio
16
„Modified“ Non-dominated Sorting
• Build „aggregated“ Fronts (Set of Pareto-Fronts), that are not dominated by remaining Pareto Fronts.
• Diversity sorting Criteria: Contribution of Pareto-Front to aggregated Front in form of length.
1. agg. Front2. agg. Front3. agg. Front
V(x)
E(x)
17
Calculation of intersection- and jump-points
• Basic Idea– Each Pareto Front is a set of segments– Segment := Part of Pareto-Front, which starts and ends at two
neighboured Corner-Portfolios.– Start with segment that contains Corner-Portfolio with highest
expected return.– Run through all segments until segment with lowest return has been
reached– Check at each segment if there is an intersection or a jump to another
segment
• Every segment defines intervals on return and variance axis.
E(x)
V(x)
18
Variance and Return within a Segment
E(x)
V(x)
19
Dominated Area
Calculation of jump-points
• Two cases where jumps are possible:1. Another Pareto-Front starts within the return-interval defined by the
current segment.2. The current segment is the most left one: jump to next best Pareto-
Front. Further Pareto-Fronts can only be counted to aggregated Front if there is no domination by variance of best-known portfolio
E(x)
V(x)
E(x)
V(x)
20
• First Idea:– Intersection with other segments is only possible, if intervals on return-
axis overlap.
Calculation of intersection-points (1)
V(x)
E(x)
21
Calculation of intersection-points(2)
• We need to check if return and variance of two segments are equal:
• Subistitute . Possible intersection is solution of quadratic equation depending on .
• depends on the position of the two segments.• Better alternative: construct artificial segments, that have equal
return-intervals.
E(x)
V(x)
22
agg Front 3
Update of Population
• Similar to NSGA-2
oldpop
off-spring
Modifiednon-domiated
sorting
agg. Front 1
agg. Front 2
agg.Front 3
…
agg. Front k
Diversity sorting
agg. Front 1
agg. Front 2
Form new offsprings
23
24
Literature
• Streichert, Ulmer und Zell: „Evalutating a Hybrid Encoding and Three Crossover Operators on the Constrained Portfolio Selection Problem“
• Streichert, Ulmer und Zell: „Comparing Discrete and Continuos Genotypes on the Constrained Portfolio Selection Problem“
• Streichert, Ulmer und Zell: „Evolutionary Algorithms and the Cardinality Constrained Portfolio Optimization Problem“
• Derigs und Nickel: „Meta-heuristic based decision support for portfolio optimization with a case study on tracking error minimization in passive portfolio management“