PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

We investigate the use of quantum computers for building a portfolio out of a universe of U.S. listed, liquid equities that contains an optimal set of stocks. Starting from historical market data, we look at various problem formulations on the D-Wave Systems Inc. D-Wave 2000Q(TM) System (hereafter called DWave) to find the optimal risk vs return portfolio; an optimized portfolio based on the Markowitz formulation and the Sharpe ratio, a simplified Chicago Quantum Ratio (CQR), then a new Chicago Quantum Net Score (CQNS). We approach this first classically, then by our new method on DWave. Our results show that practitioners can use a DWave to select attractive portfolios out of 40 U.S. liquid equities.
Content may be subject to copyright.
Portfolio Optimization of 40 Stocks Using
DWaves Quantum Annealer
Chicago Quantum
email the authors
July 6, 2020
Contents
1 Introduction 2
2 Validity of the Formulation 3
3 Classical Methods 3
3.1 BruteForce.............................. 4
3.2 GeneticAlgorithm .......................... 4
3.3 RandomSampling .......................... 5
3.4 HeuristicApproach.......................... 5
3.5 Simulated Annealer as a Monte Carlo . . . . . . . . . . . . . . . 5
4 Using an Annealing Quantum Computer 5
4.1 The Optimal Portfolio . . . . . . . . . . . . . . . . . . . . . . . . 5
4.2 Developing the QUBO to Number of Assets in a Portfolio . . . . 6
4.3 Embedding, Scaling, and Hardware Considerations . . . . . . . . 7
4.4 Affine Transformations of the QUBO . . . . . . . . . . . . . . . . 7
4.5 Visualization ............................. 8
5 Results 9
6 Discussion and Conclusion 11
7 Thank You 16
Abstract
We investigate the use of quantum computers for building a portfolio
out of a universe of U.S. listed, liquid equities that contains an optimal set
Jeffrey Cohen, Alex Khan, Clark Alexander
1
arXiv:2007.01430v1 [q-fin.GN] 2 Jul 2020
of stocks. Starting from historical market data, we look at various prob-
lem formulations on the D-Wave Systems Inc. D-Wave 2000QTM System
(hereafter called DWave) to find the optimal risk vs return portfolio; an
optimized portfolio based on the Markowitz formulation and the Sharpe
ratio, a simplified Chicago quantum ratio (CQR), then a new Chicago
quantum net score (CQNS). We approach this first classically, then by
our new method on DWave. Our results show that practitioners can use
a DWave to select attractive portfolios out of 40 U.S. liquid equities.
1 Introduction
The challenge we approach in financial portfolio optimization is to maximize
expected returns while minimizing variability of expected returns, or risk. This
is a buy and hold strategy and not a mid or high frequency trading strategy. It
relies on previous period risk, in our case one year of daily adjusted close data,
and the underlying variability and relationships of equities. We believe investors
can improve their chances by selecting the right combination of stocks.
Among the major challenges in financial portfolio optimization is “how does
an investor balance long term investments between expected return and volatil-
ity?” In this work we tackle this question from a variety of methods.This problem
is particularly well suited for an annealing solution, either classical simulated
thermal annealing, or quantum annealing since we wish to consider Nequities in
which each equity may be included in a portfolio or not. This yields exactly 2N
possibilities. For a potential list of equities as small as 40, this becomes nearly
infeasible on a workstation. When we approach the entirety of the S&P 500, we
very quickly run into a solution space which is computationally infinite. That
is, we do not have enough memory in the observable universe to run through a
brute force solution.
This work is structured as follows: In §2 we begin our exploration with the
Sharpe ratio
Sa(w) = E[RaRb] + Rb
σa
(1)
Where βis the ratio of Covariance of a portfolio with the market over the
variance of the entire market [3], Rais the return of the collection of assets,
Rbis the risk free return, and σais the standard deviation of the collection of
assets, and wis a vector of weights for assets in our portfolio.
We can also see the Sharpe ratio in matrix form as
Sa(w) = E[RaRb] + Rb
[wtCovijw]1/2(2)
From here we develop the Chicago Quantum Ratio (CQR)
CQRa(w) = w·Covim
σa
(3)
2
where Covim is the covariance of the ith asset against the entire market. This is
a slight improvement over the Sharpe ratio in terms of computation as we need
not consider nominal assets. Risk free investments have a near zero covariance
with the entire market.
We can also reformulate CQR in matrix form as
CQRa(w) = w·Covim
[wtCovijw]1/2(4)
We explore these formulations by a variety of classical methods which one
will find in §3. Both formulations are ratios and thus neither is properly suitable
for a quantum annealing solution, as DWave requires a linear quadratic form.
We attempt to rectify this by exploring
ln(Sa) = ln(E[RaRb]) ln(σa) (5)
This, however causes a different set of mathematical problems in formulating
a consistent quadratic form. Finally we settle on the Chicago Quantum Net
Score (CQNS) which is given by
CQNS(w;α) = V ar(Rw)E[Rw]2+α(6)
Where Rwis a weighted portfolio and αRIn most experiments we choose
an equal weighting i.e. wi= 1/n where nis the number of assets included,
and we choose αnear 1. These are not requirements, but they do make the
computations on DWave slightly easier. There is a wide open question as to
finding optimal weighting and optimal α.
We explain how to formulate a quadratic form for use on DWave in §4.
Finally in §5 and 6 we give our results visually and mathematically, and
discuss our future work.
2 Validity of the Formulation
In its current capacity DWave solves problems which are formulated in terms
of an Ising model. Thus our practical challenge is to provide for Dwave an ac-
ceptable model on which it may begin its computations. Consider the following
image 1
Our formulation has a propensity toward conservative side in investment
terms, however, is also demonstrates that the present formulations are near the
efficient frontier of investment portfolios. Thus from an empirical perspective
this formulation passes muster. We develop the method in more detail in §4.
3 Classical Methods
In this section we wish to give various formulations which we run on digital
computers. These methods are meant as benchmarking measures so that we
3
Figure 1: Comparison of CQR and CQNS scores against the Sharpe Ratio
may check whether the computations from the annealing quantum computer
are legitimate or if the annealing computer lands on local minima which are not
particularly deep.
3.1 Brute Force
For a smaller asset universe we are able to simply loop through all binary solu-
tions. If given enough time one can brute force roughly 40 assets. This however
is best approached by an in-place in-time algorithm. If one attempts to build
all 2Nportfolios first and simply loop through a list, one will run out of mem-
ory. An in-place algorithm eliminates the memory excess. An in-time algorithm
allows us to write out solutions in case of interruption. With brute force meth-
ods we can explicitly know the minimum possible energy level and thus verify
whether our formulation for an annealing quantum computer is valid.
3.2 Genetic Algorithm
Our genetic algorithm solution gets to a local minimum deeper than our Monte
Carlo method with 950M samples, and does so very quickly. Our difficulty is in
tuning the parameters for number of evolutionary steps, probability of elitism,
4
and size of initial population. Even with essentially random guesses at these
parameters, our genetic algorithm reaches a low enough energy level so that we
can determine whether the quantum annealing solutions are legitimate.
3.3 Random Sampling
As mentioned earlier §3.1 a 40 asset portfolio is slightly more than a workstation
can handle without an in-place algorithm. Thus we randomly sample as much
as we can. We are able to sample roughly 229 portfolios of a potential 240 This
means most of our effort is spent around portfolios of size 40 ±40. Percentage
wise this doesn’t cover much of the entire spectrum, but we approach 0.4% of
the mid sized portfolio. Specifically by Stirlong’s approximation we know.
P(|X|= 20) = 1
240 40
201
20π0.12
On the other hand 229/240 0.0005. So we get about an 8 fold lift around the
middle portfolios.
3.4 Heuristic Approach
After running a number of the previous classical methodologies we notice that
certain stocks appear most often within the best performing portfolios. We
name these stocks “All stars.” Similarly we notice several stocks which appear
most often in the worst performing portfolios. We name these “Dog stars.” The
heuristic approach is to attempt building portfolios of mostly All stars with the
addition of a few extra items. This works well as a seeding algorithm for other
probabilistic methods and will inform our approach when we attempt to solve
the portfolio optimization problem on quantum circuit and trapped ion models.
3.5 Simulated Annealer as a Monte Carlo
The original test of our problem comes in the form of a simulated (thermal)
annealing solution. Using statistics of random matrices we are able to tune the
parameters of our simulated annealing solution to deliver very deep local min-
ima. Additionally, this style solution only covers minimizing risk in a portfolio.
Based on the cooling rate of the thermal annealing and the number of attempts
per solution we use simple statistics from sampling theory to provide a measure
of goodness.
4 Using an Annealing Quantum Computer
4.1 The Optimal Portfolio
The optimal portfolio in our case is one which maximizes the Sharpe ratio.
However, as presented the Sharpe ratio of a portfolio is not computable as a
5
QUBO. The main thrust of this research is, in fact, how to formulate a QUBO
which, when presented to DWave produces similar results to the classical Sharpe
ratio. Consider the following, The Sharpe ratio is defined above in 1
Sa=βE[RaRb] + Rb
σa
The numerator can be expressed as a simple dot product, i.e.
Xµiwi
where µiand wiare the expected return and the relative weight of the ith
asset, respectively. The denominator can be expressed as the square root of a
quadratic form, i.e.
1
|U|2Xviqi+ 2 X
i<j
covij qiqj
1/2
Where qi∈ {0,1}is a binary classifying whether the ith asset is in the
portfolio or not, viis the variance, and covij is the covariance term between
asset iand j. One will immediately recognize this as σaas in the initial formula.
One will also recognize that the Sharpe ratio is not a proper quadratic form, and
thus not suitable for DWave in its current iteration. We find that the Chicago
Quantum Net Score (CQNS) solves this problem and can be presented as a
quadratic form.
4.2 Developing the QUBO to Number of Assets in a Port-
folio
Consider a universe Uof Nassets. When dealing with a single asset portfolio,
we only consider the linear terms in a QUBO. In particular when we have a
lower triangular matrix or a zero diagonal matrix, products of the form
et
iQei= 0
Thus we pick off only the linear terms. In this case, we concisely model the
inverse Sharpe Ratio on qubits and use a penalty on the couplers. DWave finds
one-asset portfolios with the highest ratios.
Moving to two or more assets we have substantially more work to do. Look-
ing at a single asset, there are no covariance terms to deal with, and we can
embed the inverse Sharpe ratio directly onto the qubits. We create a unique
QUBO for each size portfolio evaluated {2, . . . , N}by applying the weights di-
rectly to the matrix, so qiand qjcan remain binary. We divide the linear terms
by N(Pwi= 1) and apply the linear affine transformation. We divide the
variance terms (the diagonal entries) by N2(N1) to avoid duplication, and
divide the covariance terms (off diagonal entries) by N2to avoid duplication.
6
We then apply the quadratic affine transformation. Then we assemble the ma-
trix and reverse the sign on the linear terms. Finally, we apply a scale factor
(1,1) to the QUBO and write it into our N×N×Nmatrix for processing by
DWave
4.3 Embedding, Scaling, and Hardware Considerations
We embed the CQNS on DWave by writing the expected returns onto the linear
terms, both variance (diagonal) and covariance (off diagonal) onto the quadratic
terms. From here the DWave inspector shows how the system encodes and
embeds assets onto physical qubits. An attempt at changing the formulation
by manually embedding terms to respect a reordering of assets does not yield
substantial improvements in performance thus we use DWaves automated em-
bedding functions.
As we increase our asset size, we see that for a fully connected QUBO DWave
requires multiple qubits in chains to leverage the available connections in other
groups of the chimera structure. Increasing our portfolio size above 40 would
result in increasing qubit counts utilized to support multiple chains, and the
potential for increased chain breaks. We see consistent results with 40 assets
when we tune the “Chain Strength and scale factors.
We attempt to reduce our resource cost by removing links between assets that
are thought to be insignificant due to low correlation values in our calculations.
However, we create inaccurate results in this formulation and from this point
we shall avoid this method of reducing our resource cost.
We adjust DWaves chain strength parameter to see adequate results with low
volume of chain breaks. DWave defaults to a chain strength of 1, and we found
we could reduce chain breaks by setting chain strength to values as high as 15.
However, a chain strength of 1 provides more valid answers to our particular
QUBO.
We control the qubit value scaling to avoid unequal scaling of linear and
quadratic terms. Dwave’s native scaling, if left untouched would reduce the
accuracy of our scores. We scale the values we send to DWave within the
QUBO. The process of scaling naturally moves us toward scaling our values by
a hyperbolic tangent. In particular, we find the original scale of our values has
some very small values owing to covariance matrices having a zero eigenvalue.
The hyperbolic tangent scales our values to the range (1,1) we cutoff at ±0.99.
The original scaling produced results which are difficult to read and slightly
inconsistent. The newer scaling gives more reliable and consistent results.
4.4 Affine Transformations of the QUBO
When exploring portfolios of different sizes, we present a different matrix to
Dwave for each desired size of portfolio. We add a penalty for exploring portfo-
lios of different sizes, while maintaining accurate values for the desired portfolio
size. The intuition for this follows closely from converting a QUBO into an
Ising model. In order to convert a QUBO into an Ising model we consider the
7
transformation on the binary vector x:
z= 2x1
This transforms xtQx into ztJz +c·z+kwhere cis a vector of matching
length and kis a constant which we can remove from consideration. Since we’re
only looking for the location of the lowest energy level in zcoordinates, we
convert back to xand find the actual lowest energy. Thus our intution leads us
to consider affine transformations in x:
z=ax +b(7)
=xtQx
=ztJz +c·z+k(8)
where J=Q/a2,c=2J·band again kis a constant about which we
care none. Our goal at this point is to find a shift which we can apply to the
quadratic form. This corresponds to a translation and is thus closely related
to the term babove. In our formulation we do not use balance although our
mathematics takes it into consideration. In this case balance will correspond to
a boost in our coordinate system.
In order to make things easier from the point of view of computation we do
not give explicit formulations of shift and balance in terms of aand babove,
but rather explain what we actually compute.
Definition 1. Given a universe Uof assets from which to choose, we define the
shift factor snfor exactly n > 1 assets from Uas
sn=2gnm
|U| (9)
where gis the best score derived from a classical simulation in this case a genetic
algorithm, 1.5<m<20 is a multiplier which we derive empirically.
Our multiplier mis generally around 5. Intuitively, errors can be multiplica-
tive and so we consider values close the geometric average of 1.5 and 20 i.e.
30 5.4
As mentioned we skip the algebraic coordinate transformation and simply
add our shift factor to both linear and quadratic terms, but we do so as follows:
1. to linear terms add sn/n that is gm/|U|
2. to quadratic terms add 2sn/(n1)
4.5 Visualization
Visualization of the energy landscape is critical in learning how DWave finds a
solution. It also aids our understanding of how matrix transformations adjust
the landscape to improve the probability of sampling “correct” asset values. For
8
Figure 2: CQNS scores raw vs computed with a shifted matrix
example, we can place a penalty on smaller portfolios by adding a shift term
while subtracting a shift term places a penalty on larger portfolios. Consider
the following image:
The values in both of these graphs are computed energy values and the x-axis
is the set of assets sorted by number of assets in a portfolio. We see an unevenly
shifted matrix where we have chosen exactly half the possible assets. We achieve
a nice “U”-shaped curve where the minimal values are clustered around |U|/2
assets. Furthermore, around |U|/2 our shifted matrix gives us lower values than
the raw CQNS. We repeat this for each number of assets n > 1, while holding
the raw CQNS scores for the desired portfolio size constant. We ‘tilt the curve
toward or away from smaller portfolios, with the opposite impact on larger ones,
to ensure DWave finds the desired portfolio sizes in that QUBO.
5 Results
Our experimental workflow is as follows:
1. Download 1-year of daily market data for a specific set of N assets and
indices.
·Current as of that moment
·Hold that data for all experiments
2. Calculate covariance of each asset with the market, and β,[3] based on log
returns
3. Calculate covariance terms between assets
9
4. Calculate underlying and summary values, including Sharpe Ratio and
Chicago Quantum Net Score, for an all asset portfolio (i.e. hold all 40
assets for an equal investment amount)
5. Derive a QUBO for each portfolio size (2 to |U|).
·Visualize minimum CQNS values on multiple QUBO matrices.
·Shift each QUBO to increase likelihood of choosing a portfolio with
fixed number of assets.
6. Run a classical probabilistic algorithm, in our case a genetic algorithm, to
see one “best” portfolio and its values.
7. Execute DWave using appropriate range of portfolio sizes.
8. Use the Dwave results the seed the genetic algorithm
9. Compare values to classical methods.
The following figures 34 give some idea of how well the Quantum computer
performs using the CQNS against the Sharpe ratio. We see that in this sample,
DWave approaches the efficient frontier in a few cases (highest return for that
level of risk). Most points achieve relative parity with random sampling results,
and in some cases DWave suggests lower performing portfolios. These results
vary by sample, sample size, and market conditions. One will also notice that
toward the lower left of the efficient frontier, there is a higher density of solutions
which shows us that the CQNS formulation lands on the efficient frontier, but
is somewhat more conservative. In reality, the CQNS tends to favor portfolios
with lower risk.
We further give results of the CQNS by method. One will notice that DWave
performs well, in fact obtaining better results than Monte Carlo methods, but
underperforming the genetic algorithms. Using DWave as a seed guarantees the
genetic algorithm will perform better, as we disallow anything “worse” than
the seed to propagate through generations of solutions. Interestingly, the two
genetic algorithms give the same answers. We were lucky, however to be able
to run the genetic algorithms through many generations. We expect that if our
universe were to have 1024 assets the genetic algorithm with DWave seeds
would be the top performer, with the simple genetic algorithm performing close
to DWave.
We see that DWave outperforms classical random sampling on average, for
all portfolios n∈ {325,27,28,33}, which is where most of our portfolios were
run. This shows that DWave is not picking randomly or average solutions, but
good ones. The under performance for larger portfolios gives us food for thought
in future experiments.
10
Figure 3: DWave solutions; expected return vs standard deviation. We also plot
Expected return vs Market Momentum, which is a covariance with the market
without adjusting for nominal returns as in the Sharpe Ratio.
(a) CQR vs DWave (b) CQNS vs DWave
Figure 4: Some plots of Classical vs Quantum annealer computations
6 Discussion and Conclusion
Positive Semi-Definite Considerations Practitioners of numpy will know
well that numpy is prone to rounding errors. In particular we find that numpy
computes covariance matrices with slightly negative eigenvalues ∼ −1e8.
While this is not a particularly large negative value, it does open the possibility
of a “minimum risk portfolio” by having a negative overall variance, then the
computed Sharpe ratio will be several orders of magnitude too large. In order to
mitigate this we first test our covariance with a Cholesky decomposition.Second
we can simply modify our matrix by computing the standard eigendecomposition
11
(a) CQNS by Method (b) Completion Time by Method
Figure 5: Some plots of Classical vs Quantum annealer computations
Figure 6: CQNS by asset size
and setting all eigenvalues below some absolute threshold at exactly zero.
Considerations of Weighting Assets In order to maintain a weighting of
assets which sums to 1, we restrict from negative values. In principle one can
short assets, but designing a quadratic form to account for this is a separate
problem. In particular if we short one asset then the sum of positive investments
will be greater than one and incurs questions about over and under leveraging.
This is out of the scope of our present research. The second problem is that we
will not have a QUBO as we must consider values {−1,0,1}. This requires a
tertiary optimizer, not a binary one. For this research we have chosen an even
positive weighting of assets to reduce computation space. When we have |U|
assets, our search space with even weightings is 2|U| . Were we to allow a con-
tinuous weighting, we would have to approach this with a different optimization
scheme. If we were to allow assets to have a discrete weighting, e.g. 0 to 1 by
0.01, our search space becomes 101|U| which is approximately 7 order of magni-
tude larger. This brings our ability to search for an optimal portfolio down to
only 32. We demonstrated that a 32 asset portfolio can be optimized reasonably
well with classical methods. Specifically, we have enough solution space to line
up eigenvectors much more easily.
12
Selecting our Economic Values We use three market equity indices to
derive our one-year market returns (Wilshire 5000, S&P 500, Russell 2000),
along with the average return over the past year of 13 week US Treasury Bills.
We apply floors to each index to avoid negative market returns. We remove
stocks with β < 0 or β > 10, and those without continuous trade data over the
past year i.e. 253 days, in order to avoid market anomalies. We select one-year
data because we lose too much variability with five-year historical market data
2/3. We believe this is due to the market reverting to means.
Model adjustments during time of market turbulence: Our model for stock
selection requires mathematical adjustment in times of market declines when our
indices move into negative territory over a year or when interest rates approach
or drop below zero. This impacts the signs of our scoring models.
Discussion of Quantum Advantage We optimize a reasonably sized port-
folio using DWaves 2,041 qubit quantum annealer through a repeatable research
and business process. We pick 40 assets, which creates a solution space of 240,
or 1.1 trillion portfolios from which to select.
As practitioners, our research indicates potential for quantum advantage at
a higher number of assets. At lower asset levels there are efficient classical
algorithms that do better. We cannot solve this problem using brute force on
our equipment at 40 assets. We find an equivalent portfolio with a random
sampling of 500M portfolios that takes hours to run. One portfolio with 3 out
of 40 assets appears optimal via a genetic algorithm seeded with 1,028 random
portfolios. We improve on that timing by seeding the genetic algorithm with
the DWave solutions.
If we repeat the process of 36 experiments with 60 assets, we expect we
might beat the genetic algorithm...but we would also use a classical simulated
annealer which could outperform the genetic algorithm...so the race for quantum
advantage continues.
Next Steps In the future we intend to evaluate reverse annealing, simulated
annealing (both thermal and quantum), use of the DWave Hybrid solver, and to
optimize larger and more diverse portfolios. We also intend to further optimize
the DWave runs and the QUBO build process.
From an economics perspective, we intend to add different types of financial
assets, including bonds, commodities, real estate investment trusts, and cur-
rencies (including Bitcoin - USD). We intend to evaluate additional economic
factors such as financial health, growth, dividend payouts, and liquidity, which
today we include implicitly.
Finally, we saw unique behavior when actual market returns over one year
became negative (which it did during our research), and when individual stocks
behaved erratically (e.g., β < 0 and β > 5.0). We intend to explore how these
impact the portfolio optimization solutions from the Chicago Quantum Net
Score.
13
References
[1] N.Bekkers, R.Doeswijk, T.Lam, Strategic Asset Allocation: Determining
the Optimal Portfolio with Ten Asset Classes https://papers.ssrn.com/
sol3/papers.cfm?abstract_id=1368689 Oct 20 09
[2] StackExchange, Comparing Portfolio Volatility with Index Volatil-
ity seems a wrong method?, Jan 2015, AltTabsen, QuantK,
and Ric, https://quant.stackexchange.com/questions/16223/
comparing-portfolio-volatility-with-index-volatility-seems-a-wrong-method
[3] BETA, Investopedia https://www.investopedia.com/terms/b/beta.
asp
[4] G. Brinson, L.R. Hood, G. Beebower, Determinants of Portfolio Perfor-
mance, June 1986, Financial Analysts Journal 42(4):39-44, DOI: 10.2469/
faj.v42.n4.39
[5] G. Brinson, B. Singer, G. Beebower, Determinants of Portfolio Performance
II: An Update, May/June 1991, Financial Analysts Journal pp.40
[6] Ahmed Fasih, StackOverflow: Python: convert matrix to posi-
tive semi-definite, https://stackoverflow.com/questions/43238173/
python-convert-matrix-to-positive-semi-definite/43244194
[7] V. Divakar, Calculating the Covariance Matrix and Portfolio
Variance, December 27, 2018, https://blog.quantinsti.com/
calculating-covariance-matrix-portfolio-variance
[8] DWave Systems Documentation,https://www.docs.dwavesys.com
[9] DWave Systems Application Development Resources, https://cloud.
dwavesys.com/leap/resources
[10] Efficient Frontier at Wikipedia https://en.wikipedia.org/wiki/
Efficient_frontierEfficientFrontierWiki
[11] Harry Markowitz, Portfolio Selection, The Journal of Finance, 7,1 (1952),
77-91
[12] Michael Marzec. Portfolio Optimization: Applications in Quantum Com-
puting. SSRN Electronic Journal, 2013 https://papers.ssrn.com/sol3/
papers.cfm?abstract_id=2278729
[13] Marcos Lopez de Prado, Ph.D., Machine Learning Asset Allocation, https:
//ssrn.com/abstract=3469964, Advanced in Financial Machine Learn-
ing, ORIE 5256
[14] Nada Elsokkary, Faisal Shah Khan, Davide La Torre, Travis S. Humble,
Joel Gottlieb. Financial Portfolio Management using D-Waves Quantum
Optimizer: The Case of Abu Dhabi Securities Exchange. IEEE High Per-
formance Extreme Computing Conference, 2017 (edited)
14
[15] Bernard Pfaff, Financial Risk Modelling and Portfolio Optimization with
R, second edition, Wiley, 2016. ISBN : 9781119119661
[16] Michael Jensen, The Pricing of Capital Assets and the Evaluation of In-
vestment Portfolios, The Journal of Business, Vol. 42, No.2, (Apr., 1969),
pp. 167 - 247
[17] An Introduction to Information Geometry, https://arxiv.org/pdf/
1808.08271.pdf
[18] Ors, Romn, Samuel Mugel, and Enrique Lizaso. Quantum comput-
ing for finance: overview and prospects. Reviews in Physics 4 (2019)
(arXiv:1807.03890 [quant-ph])
[19] ReSolve Asset Management Whitepaper, A General Framework for Port-
folio Choice, 27 pages, undated
[20] M. Rubenstein, Markowitz’s Portfolio Selection: a Fifty-Year Ret-
rospective, The Journal of Finance, Vol LVII, Num 3, June 2002,
DOI:10.1.1.404.4279
[21] Sharpe Ratio at Wikipedia https://en.wikipedia.org/wiki/Sharpe_
ratio
[22] William F. Sharpe, Stanford University, The Sharpe Ratio, The Journal of
Portfolio Management, Fall 1994 http://web.stanford.edu/~wfsharpe/
art/sr/sr.htm
[23] Standard & Poor, formerly CapitalIQ company database, Highland Park
Public Library access to data screening tools, https://hplibrary.org/
databases-5317
[24] William F. Sharpe, Mutual Fund Performance, The Journal of B business,
Vol 39, No.1, Part 2, Supplement on Security Prices. (Jan 1966), pp. 119 -
138 http://www.jstor.org/stable/2351741
[25] Scot Stockton, What’s The Difference Between 45 and
28 Percent Return? The Efficient Frontier, August 20,
2018, Seeking Alpha, https://seekingalpha.com/article/
4200744-difference-45-return-and-28-efficient-frontier
[26] D.Venturelli, A.Kondratyev, Reverse Quantum Annealing Approach to
Portfolio Optimization Problems, arXiv:1810.08584, 2018
[27] X.Ye, L.Ning, J.Yang, D.Yao, Y.Hong, Research on three different Portfo-
lio Models with singular Covariance Matrix IOSR Journal of Mathematics
(IOSR-JM) Volume 14, Issue 5 Ver. II (Sep - Oct 2018), PP 33-39
[28] YFinance python module, maintained by Ran Aroussi, Thank you https:
//pypi.org/project/yfinance/
[29] Yahoo Finance provides all historical data used in these experiments,
Thank you, https://finance.yahoo.com/
15
7 Thank You
We acknowledge and thank the writers, maintainers and community contribu-
tors for Python, Numpy, Pandas, Matplotlib, Scipy, and DWave Ocean, Julia,
and R. We also thank DWave Systems, Google, Slack, Anaconda, and Jupyter
for use of their tools in this research effort.
16
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
We discuss how quantum computation can be applied to financial problems, providing an overview of current approaches and potential prospects. We review quantum optimization algorithms, and expose how quantum annealers can be used to optimize portfolios, find arbitrage opportunities, and perform credit scoring. We also discuss deep-learning in finance, and suggestions to improve these methods through quantum machine learning. Finally, we consider quantum amplitude estimation, and how it can result in a quantum speed-up for Monte Carlo sampling. This has direct applications to many current financial methods, including pricing of derivatives and risk analysis. Perspectives are also discussed.
Book
This book introduces the latest techniques advocated for measuring financial market risk and portfolio optimization, and provides a plethora of R code examples that enable the reader to replicate the results featured throughout the book. This edition has been extensively revised to include new topics on risk surfaces and probabilistic utility optimization as well as an extended introduction to R language. Financial Risk Modelling and Portfolio Optimization with R: Demonstrates techniques in modelling financial risks and applying portfolio optimization techniques as well as recent advances in the field. Introduces stylized facts, loss function and risk measures, conditional and unconditional modelling of risk; extreme value theory, generalized hyperbolic distribution, volatility modelling and concepts for capturing dependencies. Explores portfolio risk concepts and optimization with risk constraints. Is accompanied by a supporting website featuring examples and case studies in R. Includes updated list of R packages for enabling the reader to replicate the results in the book. Graduate and postgraduate students in finance, economics, risk management as well as practitioners in finance and portfolio optimization will find this book beneficial. It also serves well as an accompanying text in computer-lab classes and is therefore suitable for self-study.
Article
Portfolio optimization in a quantum computing paradigm is explored. The D-Wave adiabatic quantum computation optimization system is used to determine an optimal portfolio of stocks using binary selection. The stock returns, variances and covariances are modeled in the graph-theoretic maximum independent set (MIS) and weighted maximum independent set (WMIS) structures. These structures are mapped into the Ising model representation of the underlying D-Wave optimizer. The results show different stock selections over a range of predetermined risk thresholds and underlying models. This implementation and following discussion provides a practitioner’s view of what might be accomplished in this framework. The particular models used in the implementations have restricted appeal but do link the financial engineering domain to the quantum computing optimization domain. Further research on model enhancements or different model structures needs to be undertaken to improve its usefulness in comparison to the current industrial domain.
Article
THIS YEAR MARKS the fiftieth anniversary of the publication of Harry Markowitz's landmark paper, "Portfolio Selection," which appeared in the March 1952 issue of the Journal of Finance. With the hindsight of many years, we can see that this was the moment of the birth of modern financial economics. Although the baby had a healthy delivery, it had to grow into its teenage years before a hint of its full promise became apparent. What has always impressed me most about Markowitz's 1952 paper is that it seemed to come out of nowhere. Compared to the work of his 1990 co-Nobel Prize winners (Sharpe primarily for his paper on the capital asset pricing model and Miller for his paper on capital structure), Markowitz's paper seems to have more of this flavor. In 1676, Sir Isaac Newton wrote his friend Robert Hooke, "If I have seen further it is by standing on the shoulders of giants" (Newton (1959)) and that is true of Markowitz as well, but, like Newton, he certainly saw a long distance given the height of those shoulders. Markowitz was hardly the first to consider the desirability of diversification. Daniel Bernoulli in his famous 1738 article about the St. Petersburg
Article
This study explores which asset classes add value to a traditional portfolio of stocks, bonds and cash. Next, we determine the optimal weights of all asset classes in the optimal portfolio. This study adds to the literature by distinguishing ten different investment categories simultaneously in a mean-variance analysis as well as a market portfolio approach. We also demonstrate how to combine these two methods. Our results suggest that real estate, commodities and high yield add most value to the traditional asset mix. A study with such a broad coverage of asset classes has not been conducted before, not in the context of determining capital market expectations and performing a mean-variance analysis, neither in assessing the global market portfolio.
Article
The main purpose of this study is the development of a model for evaluating the performance of portfolios of risky assets taking into account the effects of differential risk on required returns. The portfolio evaluation model developed here incorporates these risk aspects explicitly by utilizing and extending recent theoretical results by Sharpe (1964) and Lintner (1965) on the pricing of capital assets under uncertainty. Given these results, a measure of portfolio performance (which measures only a manager's ability to forecast security prices) is defined as the difference between the actual returns on a portfolio in any particular holding period and the expected returns on that portfolio conditional on the riskless rate, its level of systematic risk, and the actual returns on the market portfolio. Criteria for judging a portfolio's performance to be neutral, superior, or inferior are established. A measure of a portfolio's efficiency is also derived, and the criteria for judging a portfolio to be efficient, superefficient, or inefficient are defined. I also show that it is strictly impossible to define a measure of efficiency solely in terms of ex post observable variables. I define two forms of the efficient market hypothesis, the weak form and the strong form (following terminology introduced by Harry Roberts, who used these terms in an unpublished speech entitled Clinical vs. Statistical Forecasts of Security Prices, given at the Seminar on the Analysis of Security Prices sponsored by the Center for Research in Security Prices at the Univ. of Chicago, May 1967. One can define a weakly efficient market in the following sense: Consider the arrival in the market of a new piece of information concerning the value of a security. A weakly efficient market is a market in which it may take time to evaluate this information with regard to its implications for the value of the security. Once this evaluation is complete, however, the price of the security immediately adjusts (in an unbiased fashion) to the new value implied by the information. In such a weakly efficient market, the past price series of a security will contain no information not already impounded in the current price. In such a market, forecasting techniques which use only the sequence of past prices to forecast future prices are doomed to failure. The best forecast of future price is merely the present price plus the normal expected return over the period. The available evidence suggests that it is highly unlikely that an investor or portfolio manager will be able to use the past history of stock prices alone (and hence mechanical trading rules based on these prices) to increase his profits. However, the conclusion that stock prices follow the weak form of the efficient market hypothesis allows for an investor to increase his profits by improving his ability to predict and evaluate the consequences of future events affecting stock prices. This brings us to the strong form definition of an efficient market, that is, one in which all past information available up to time t is impounded in the current price. If security prices conform to the strong form of the hypothesis, no analyst will be able to earn above-average returns by attempting to predict future prices on the basis of past information. The only individual able to earn superior returns will be that person who occasionally is the first to acquire a new piece of information not generally available to others in the market. But as Roll (1968) argues, in attempting to act immediately on this information, this individual will insure that the effects of this new information are quickly impounded in the security's price. Furthermore, if new information of this type arises randomly, no individual will be able to assure himself of systematic receipt of such information. Therefore, while an individual may occasionally realize such windfall returns, he will be unable to earn them systematically through time. While the weak form of the hypothesis is well substantiated by empirical evidence, the strong form of the hypothesis has not as yet been subjected to extensive empirical tests. The model developed in this paper allows us to submit the strong form of the hypothesis to such an empirical test - at least to the extent that its implications are manifested in the success or failure of one particular class of extremely well-endowed security analysts. I use the portfolio evaluation model developed here to examine the results achieved by the managers of 115 open end mutual funds. The main conclusions are: 1) The observed historical patterns of systematic risk and return for the mutual funds in the sample are consistent with the joint hypothesis that the capital asset pricing model is valid and that the mutual fund managers on average are unable to forecast future security prices. 2) If we assume that the capital asset pricing model is valid, then the empirical estimates of fund performance indicate that the fund portfolios were inferior after deduction of all management expenses and brokerage commissions generated in trading activity. When all management expenses and brokerage commissions are added back to the fund returns and the average cash balances of the funds are assumed to earn the riskless rate, the fund portfolios appeared to be just neutral. Thus, on the average the resources spent by the funds in to forecast security prices do not yield higher portfolio returns than those which could have been earned by equivalent risk portfolios selected (a) by random selection policies or (b) by combined investments in a market portfolio and government bonds. 3) I conclude that as far as these 115 mutual funds are concerned, prices of securities seem to behave according to the strong form of the efficient market hypothesis. That is, it appears that the current prices of securities completely capture the effects of all information available to these 115 mutual funds. Although these results certainly do not imply that the strong form of the hypothesis holds for all investors and for all time, they provide strong evidence in support of that hypothesis. 4) The evidence also indicates that, while the portfolios of the funds on the average are inferior and inefficient, this is due mainly to the generation of excessive expenses.
Determinants of Portfolio Performance
  • G Brinson
  • L R Hood
  • G Beebower
G. Brinson, L.R. Hood, G. Beebower, Determinants of Portfolio Performance, June 1986, Financial Analysts Journal 42(4):39-44, DOI: 10.2469/ faj.v42.n4.39
StackOverflow: Python: convert matrix to positive semi-definite
  • Ahmed Fasih
Ahmed Fasih, StackOverflow: Python: convert matrix to positive semi-definite, https://stackoverflow.com/questions/43238173/ python-convert-matrix-to-positive-semi-definite/43244194