Society of Petroleum Engineers

SPE 82012 Portfolio Optimization Techniques for the Energy Industry

Rodney W. Lessard, PhD. / Schlumberger Information Solutions

Copyright 2003, Society of Petroleum Engineers Inc. This paper was prepared for presentation at the SPE Hydrocarbon Economics and Evaluation Symposium held in Dallas, Texas, U.S.A., 5-8 April 2003. This paper was selected for presentation by an SPE Program Committee following review of information contained in an abstract submitted by the author(s). Contents of the paper, as presented, have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material, as presented, does not necessarily reflect any position of the Society of Petroleum Engineers, its officers, or members. Papers presented at SPE meetings are subject to publication review by Editorial Committees of the Society of Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper for commercial purposes without the written consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836 U.S.A., fax 01-972-952-9435.

benefit of applying portfolio theory is that the process itself provides additional insight into a decision- makers asset. An organization, with even a modest asset base size of 200 assets, is presented with more possible portfolios to pursue than hydrogen atoms in the Sun - about 1057. The Appendix outlines how these numbers are determined. These possible portfolios represent what is termed the 'search space'. Historically, a common approach to solving this problem has been the 'rank and cut' method. Using this approach, a company will rank their opportunities based on some economic metric. They then select the opportunities beginning at the top of this ranking and continue until a certain financial constraint, often a capital spending limit, has been met. Unfortunately, this approach is rarely adequate, as a typical company will measure their performance against more than one economic metric and will have many constraints. Such complexity cannot be handled through the common approach of a rank and cut. Linear Programs Linear programming is a mathematical technique used to solve constrained optimization problems that can be expressed in the following form:
N

Abstract Energy companies typically face asset investment decisions which must satisfy a number of constraints in the form of annual corporate goals and complex asset dependencies. The aim of portfolio optimization is to select a portfolio, which maximizes/minimizes one or more value or risk measures, and satisfies these constraints. In many cases, these problems can be readily expressed in the form of a set of linear equations, such that, global optimum values can be determined through the method of Linear Programming. As the complexity of corporate goals increase, non-linearities are often introduced, which require more advanced optimization techniques. In this paper we will describe a linear model for solving many optimization problems. In the case of non-linear problems we will describe the utilization of Genetic Programming as an efficient search algorithm to seek global optimum values. Finally, we will combine the two methods into a strategy for solving complex optimization problems, utilizing hybrid techniques that combine both linear and non-linear methods. Introduction The problem Exploration and Production (E & P) companies face today is that they typically have more assets to choose from than capital required to develop them. To provide long term corporate planning, companies must consider which assets to invest in, when to invest, and for some, with how much equity? The answer to these questions defines a portfolio of opportunities. Historical methods used to determine the allocation of capital towards an asset have typically involved evaluating the asset in isolation. The asset would have to stand up against certain economic tests and if it failed, the asset would not be pursued. Modern portfolio theory takes an alternate approach. The most important principle of portfolio analysis theory is that emphasis must be placed on the interaction between assets. The primary benefit of applying portfolio theory is an increase in the financial return given an acceptable level of risk. A secondary

Maximize (Minimize) such that:

N

z = Z cixi,
i=1

(1)

ZaiXi ^ b,
i=1

(2) (3)

and x > 0.

The values of a and c are constants of the solution vector x, and b is a constant of the problem. Equation (1) is often referred to as the objective function to be optimized. Equations (2,3) represent constraints on allowed values of the solution vector and any number of constraints of this form can be included. The inequality expressed in Equation (2) can be either greater than or equal to, less than or equal to or equal to. When x can take on continuous real values, Equations (1-3) can be solved using Simplex1 or Barrier2 algorithms. When x

SPE 82012

is restricted to integer values, for example 0 or 1, the techniques used to determine optimum solutions are more precisely stated as Integer Programming. When a mix of variable types is included, the problem is referred to as Mixed Integer Programming2. The models used to express portfolio optimization will generally fall under the latter category. Many commercial tools are available to solve these linear problems. For the results given in this paper we used the CPLEX solver by ILOG, which supports Simplex, Barrier and Mixed Integer Programming solvers. However, it is not the purpose of this paper to discuss these particular techniques, but rather how to express the problem of portfolio optimization into a form suitable for solving with these methods, that is, in the form of Equations (1-3). Optimizing Equity Many economic metrics for a portfolio can be expressed as linear combinations of the value of each asset proportioned by equity. For example, in many cases, the Net Present Value (NPV) of a portfolio is the sum of the equity-proportioned npv of each asset, similarity for annual values of production, p, or capital, c. As an example, we describe the optimization of a portfolio of N assets. Our example maximizes NPV and provides annual metric constraints on production, Pk, and capital, Ck, over each of the next L years: For example, xu, x12, x13 and x14 represents the equity N (4) Maximize such that

resulting production will be shifted into the future. As a result, time to invest really behaves as a completely new opportunity and thus should be modeled as a new equity variable, x,. Each variable represents the amount of equity to choose for the complete life of the asset, if selected to begin development in the assigned year.

Consider a single asset with equity variable, xij-, which may be developed in any one of the next Mi years, with each year denoted by the indexj. The contribution of this asset to the objective can be represented by:
Mi N

Maximize (Minimize )

z = /,/ ,
j=i i=i

(6)

choices of the first asset in each of the next four years. To this, we must add the constraint that only one of xi j can be chosen for each index i - afterall, they represent the same asset. In the case of integer restriction on the value of xi j, the constraint is easily written as:

Xi,j * i. =i
j

(7)

NPV = npvv
i=i

plMxt > Pk, fork = 1.L,

i=i
N

ik i *
c x

fork

=1L
(5)

i=i

x > 0,

fori = 1..N.

where npvi represents the value of Net Present Value and xi the equity interest of the ith asset. p,k, and c,k represent the production and capital for the ith asset in the kth year respectively. Thus far, we have developed a simple model of portfolio optimization that considers only the selection of assets and the equity to invest in them. However, for efficient allocation of capital and long term planning, we must also consider the time to invest. Optimizing Time to Invest Typically, the planning cycle naturally lends itself to considering the time to invest on time scales of years; therefore, we reduce the problem of time to invest into the question of which year to invest? For example, should we develop a particular asset today or two years from today? In many cases the economics for developing an asset two years from today will be different, such as the discounted value will be reduced. Contributions of individual assets to corporate constraints will also be altered, as annual capital costs and

In the case that xij- can take on real continous values, Equation (7) could not constrain the selection of only one of Xj. Equation (7) only constrains the sum to be less than or equal to unity. For example, in the continuous real value case, equity values of 0.2, 0.3 and 0.5 could be chosen and still satisfy Equation (7). When equity is restricted to integer values (0 or 1) Equation (7) works correctly. The solution to real continuous equity problem is to introduce an additional integer variable, ixi j, to the model, whose value equals one when xi j is non-zero. This integer variable does not contribute to the objective function (objective coefficients equal zero) but rather is only used to track the inclusion of an asset in a particular year. The desired behavior is achieved through the following constraints:
x

i,j - ixi,j *

for each i and j, (8)

When xij is greater than zero, the constraint forces ix, to be equal to one. The condition that only one of xij- be chosen can then be expressed as:

ixu

(9)
j=i

Returning to our previous example, Equations (4,5), the complete problem would be expressed as:
Mi N

NPV = npvi]xi]
j=i i=i

(i)

SPE 82012

where npvlj is the Net Present Value of asset i developed in year j. The complete set of constraints are written as:
Mi N

M N

ZZ
j=1 i=1

ZZ P - j !*i.j * P , fork = 1.L,

k
+

< Fk
M N

j=i i=i
Mi N
x

ZZ P.
j=1 i=1

- j+1"v i,j j=1 i=1

i,

j < Ck,

fork

= i-A

Xj
Mi j=1

- x - 0, for each i and j, (11)

j

Z ix j 1, for each i, Xj

> 0, for each i and j.

where NQ represents the sum over the subset of Q assets. Similar constraints can be expressed for at most or exactly R assets. In the case that exactly R assets must be included the inequality is replaced by equality. Many other types of dependencies can be modeled including conditional types, "If All Assets From Region Q Are Included Then Also Include Assets R". Other asset dependencies include schedule constraints, for example, "Asset Q Must Follow Two Years After Asset R, If Included". Multiple Constraint Problem

The second index on p and c is adjusted by the number of years the asset is moved into the future. Linearizing Constraints on Ratios Corporate strategies sometimes include constraints on economic metrics that are not linear functions of asset equity. However, there are a few non-linear functions that can be rewritten into linear form suitable for solving with Linear Programs. For example, the finding cost of a portfolio is expressed as the ratio of capital over production. In this case, the constraint on the finding cost, Fk, is expressed as: but, can be re-written as:
M N
C

Set of All Portfolios

of Portfolios That Satisfy Production Constraint

i,k - j+1 Xi, j

(12)
i,k - j+i i, j
x

Z Z j+i - FkP,,k-j+i) \j < o, j=1 i=1

(i

(13)

fork = 1..L, which is a linear

function of asset equity. Asset Dependencies Thus far, we have considered constraints on the contribution of
Set of Portfolios That Satisfy Capex

simultaneously satisfies all constraints. The case of a feasible problem is depicted in Figure 1 and the case of an infeasible problem is depicted in Figure 2. One of the values of Linear Figure 1: Depiction of a multiple constraint problem. The pink circle represents the set of all portfolios. The green circle represents the set of portfolios that satisfy a constraint on Capital. The blue circle represents the set of portfolios that satisfy a constraint on Production. The intersection of these two sets represents the set of portfolios that satisfy the multiple constraint problem. Programming to portfolio optimization is the speed at which the feasibility of the constraints can be determined. As companies develop a corporate strategy, they typically over- constrain the problem and are left in a state of uncertainty as to how they should adjust the problem so that it can be solved. This adjustment may be in the form of adding more assets or by adjusting the corporate constraints. However, the type of assets needed to achieve

assets to a limit of a metric: capital, production or cashflow, etc. An additional type of constraint considers inter-asset dependencies rather than the contribution of assets to metric limits. These constraints may be in the form of limits on the number of assets allowed/required from a region or by asset type. For example, the constraint that at least R from the set of Q assets be included in the portfolio, can be written as: Feasibility A linear problem is considered infeasible if no solution vector
M N Q

ZZ *,j * R.
j=1 i=1

(14)

SPE 82012

feasibility or exactly how the strategy should be adjusted is unknown. In this situation, the solution that minimizes the sum of the metric constraint failures (as depicted in Figure 2) provides sufficient insight into the problem thereby making the necessary adjustments clear.

(i)

SPE 82012

Set of Portfolios That Satisfy Production Constraint-

Multiple Constrain t Problem ?A

Set of All Portfolios 3- o

Less Than Constraint

for each metric constraint, k=1.. L.

/
03
I

IS 3S 3*

Greater Than .Constraint Metric Constraint Target

an

this

Figure 2: Depiction of infeasible multiple constraint problem. In case the set of

Set of Portfolios That Satisfy Capex Constraint Best Infeasible portfolios that satisfy each Portfolio do not intersect, i.e. there is

the constraints solution that

I
V

Minimize sum of constrai nt failures

of no satisfies both constraints. The best infeasible portfolio is the solution that finds

the balance between 3 O these sets thereby minimizing the sum of the constraint l/l O. failures. =r c Constraint failure is ai o defined as the amount by which the value of a metric for a portfolio exceeds its limit as shown in Figure 3. The value is zero when the constraint is satisfied and increases linearly when the constraint is violated. Although this function appears non-linear due to the discontinuity at the metric constraint target, it may be linearized by the introduction of a substitution variable, sk, which divides the solution into two piece-wise linear domains. Such a variable is required for each year of each goal, totalling L. Equation (15) redefines the objective and Equation (16) provides the needed constraints. Minimize z = Z
;

Figure 3: Constraint failure is a piece-wise linear function that equals zero when the constraint is satisfied and increases L L linearly when the constraint is violated. c 't Convex Hull of Feasible Portfolios o An additional analysis that can be performed using 1 optimization techniques is the determination of the convex t hull of the set of all feasible portfolios. The set of all feasible l portfolios is the set that satisfies all constraints independent of C D the objective function (depicted in Figure 1). Optimization c seeks the one feasible portfolio with the maximum or O minimum value of the objective function. However, often we O seek the optimization of the objective function at the expense of the value of another dependent economic metric. The interior region of the halo in Figure 4 depicts the distribution of all feasible portfolios plotted against both the objective function and another dependent metric.
R
Max Dependent

'

Measure

Min Dependent

Measure

Convex Hull of All Feasible Portfolios

k=1
Dependent Measure

such that:
N Mi

Z Z P j+1 i,
i,k
x

Figure 4: The interior of the halo represents the set of all feasible portfolios plotted against the objective function and another dependent metric. The boundary or halo itself is referred to as the convex hull of this set.

j Pk i=1

j=1
N Mi

ZZ i,k j+1
C

i=1

k x

(15)
i,j

j=1

h - 0.

/
(16)

SPE 82012

The outer boundary of this region defines the maximum and minimum value of the objective function for the complete range of the dependent measure. This outer boundary is known as the convex hull. The convex hull explores the trade-off between any two metrics. The classic example of this form of analysis is the Efficient Frontier. The Efficient Frontier is the collection of portfolios that provide the minimum risk for the complete range of feasible values. This region is depicted in the lower red region of Figure 4. The following process determines the convex hull for any metric pair: 1. 2. 3. Maximize the dependent metric. Minimize the dependent metric. Minimize the objective function over a number of steps between maximum and minimum dependent metric with the addition of the following constraint:
Mt N

very technique of the Simplex, Barrier and Mixed Integer Programming methods, do not lend themselves to provide many alternate solutions which may exist in the vicinity of the globally optimum solution. For these reasons we also consider the utilization of other optimization methods such as the Genetic Algorithm (GA). Genetic Algorithms, as their name implies, are based on the process of natural genetics present in everyday biological systems. It is important to note that a Genetic Algorithm is a search engine, not a solver. As discussed above, a solver, like Linear Programming, typically requires some kind assumptions on the shape or form of the solution space. A Genetic Algorithm however, is not subject to this limitation. In fact, their central theme is robustness. A well-designed Genetic Algorithm should be capable of handling large or small problems, with any set of arbitrarily complex constraints applied. At their highest level, they are basically guided random search engines. The details of Genetic Algorithms are found in Goldberg3 and their utilization in portfolio optimization is cited in Fichter4. Briefly, the iterative process uses random number generation along with sophisticated heuristics to guide it towards a set of optimal solutions as depicted in Figure 6. On each iteration, the Genetic Algorithm generates a set of portfolios. Some portfolios are generated on a purely random basis and others are determined from the last iteration's set of portfolios through a process known as crossover. The portfolios generated are ranked according to their performance against the constraints and objectives. Those that rank the highest are maintained for the next iteration (with some alterations applied in an attempt to improve upon them). This process continues until some condition for halting is reached. The result is an algorithm that searches all possible solutions (through the random portfolios) while maintaining the best. Thus all possible areas in which solutions may lie are examined.

YL
j=1

i.j*i.j

* Dep

(17)

i=1

where di j is the value of the dependent metric for the ith asset started in the jth year. Dstep is the value of the constraint added. The concept is not restricted to measures of value and risk, but can be extended to any metric pair where the optimization of one metric is obtained at the expense of another. For example, consider the optimization of net production over the range of net capital. The convex hull explores the maximum production achievable over the complete range of allocated capital.

Figure 5: An example of the convex hull where value is plotted against risk. The line shows the Efficient Frontier obtained by maximizing value for several levels of risk. The points below the line are random samples of portfolios that satisfy the constraints; in this case, we used a constraint on capital. Genetic Algorithms Aside from the advantages offered by Linear Programming techniques, there are cases when the value of economic metrics for a portfolio cannot be derived by linear equations. In addition, the

Hybrid Techniques Hybrid techniques refer to methods that sequentially apply one or more optimization techniques in an effort to improve performance or utilize the strengths of multiple techniques to solve different aspects of the same problem. For example, a constrained optimization problem may have linear constraints but a non-linear objective function. Depending on how "tight" the constraints are, a search algorithm may struggle to discover the small feasible solution space. In this case, the Linear Program may be used to

SPE 82012

solve the constraint problem, that is, provide a feasible solution which can then be used as a starting point for the sampling algorithm, thereby giving it a heasdstart. A second advantage of the Hybrid method is the ability to search the space in the vicinity of the globally optimum solution returned by a Linear Program. In very few cases the solution returned by a Linear Program will be the final portfolio that a company wishes to implement. Rather, it should be viewed as a guide to which selection of assets provide optimum performance. Searching the vicinity of the global solution provides additional learning about which assets are often included in the "good" portfolios and which assets are easily replaced. Figure 7 shows the results of sampling the space around the optimum solution (portfolio with the greatest value in this case) with a Genetic Algorithm.

that is, problems including both linear and non-linear components can be solved using Hybrid methods. Acknowledgements I would like to thank Schlumberger Information Solutions for their support throughout this research. I would also like to thank Courtney, Deanna and Suzanne for their careful reading of this manuscript. References
1. 2. 3. 4. Danzig, G.B., 1963, Linear Programming and Extensions, Princeton University Press, Princeton, NJ. Schrage, L., 1997, Optimization Modeling with LINDO, Brooks/Cole Publishing, Pacific Grove, CA. Goldberg, D.E., 1989, Genetic Algorithms in Search, Optimization and Machine Learning, Addison Wesley Press. Fichter, D.P., 2000, Annual Technical Conference of the SPE, 62970.

Appendix - Calculation of Total Number of Possible Portfolios In the simple case that assets may be either in or out of a portfolio, the total number of unique portfolio combinations would be: N portfolios = 2n assets = 2200 = 1060 portfolios. The number of hydrogen atoms in the Sun is estimated from the mass of the Sun divided by the mass of the hydrogen atom - the Sun is comprised of greater than 90% hydrogen: N hydrogen atoms = mass Sun / mass hydrogen atom = 2x10 33 g / 2x10-24 g = 1057 atoms.

Mean of Cum After-Tax Cashflow

Figure 7: A collection of portfolios provided by seeding the Genetic Algorithm with the optimum solution found by the Linear Program. Conclusions Many economic metrics for a portfolio can be expressed as linear combinations of the value of each asset proportioned by equity. In these cases, portfolio optimization is efficiently performed using the techniques of Linear Programming. In cases when economic models for a portfolio are not linear, search methods like the Genetic Algorithm can provide efficient optimization results. In addition, combined problems,