View
2
Download
0
Category
Preview:
Citation preview
ORNL is managed by UT-Battelle, LLC for the US Department of Energy
Markowitz Portfolio Optimization with a Quantum Annealer
Erica Grant, Travis Humble
Oak Ridge National Laboratory
University of Tennessee
2
Portfolio Selection
Goals:• Maximize returns
• Minimize risk
• Stay within budget
Inputs:• Uniform random
historical price data
• Budget
• Risk tolerance
Output:• A portfolio
representing a list of investments and the expected return Investment Price Expected
Return/Day1= buy0 = pass
Stock A $150 + $0.10 0Stock B $200 + $0.25 1Stock C $250 + $0.50 0
Budget: $200 Risk Tolerance: Low
3
Portfolio Selection as Binary Integer Programming
𝑥 𝜖 {0, 1}
𝑖 = asset number
𝑝+= price
𝑟+ = expected return
𝑏 = budget
max1
∑+ 𝑟+𝑥+
𝑠. 𝑡. ∑+ 𝑝+𝑥+ ≤ 𝑏
4
Unconstrained Markowitz Formulation𝑚𝑎𝑥1
𝑓(𝑥)
𝑓 𝑥 = 𝜃=>+
𝑥+𝑟++𝑥+ − 𝜃@ >+
𝑥+𝑝+𝑥+ − 𝑏@
− 𝜃A>+,B
𝑥+𝑐𝑜𝑣 ℎ+, ℎB 𝑥B
• 𝑏 = 𝑏𝑢𝑑𝑔𝑒𝑡, 𝑝+ = 𝑝𝑟𝑖𝑐𝑒, 𝑟+= 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑟𝑒𝑡𝑢𝑟𝑛, 𝑥+ 𝜖 0, 1
• Weights: 𝜃=, 𝜃@, 𝜃A ≥ 0 s.t. 𝜃= + 𝜃@ + 𝜃A = 1
• ℎ+ = vector of historical price data
• 𝑐𝑜𝑣 ℎ+, ℎB =∑NOPQ (RS,NTURS)(RV,NTRV)
WT=where m= number of price points
𝒙𝒊 = 𝟏 → 𝒃𝒖𝒚 𝒙𝒊 = 𝟎 → 𝒅𝒐𝒏c𝒕 𝒃𝒖𝒚
5
Unconstrained Markowitz Formulation𝑚𝑎𝑥1
𝑓(𝑥)
𝑓 𝑥 = 𝜃=>+
𝑥+𝑟++𝑥+ − 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ − 𝜃A>+,B
𝑥+𝑐𝑜𝑣 ℎ+, ℎB 𝑥B
• 𝑏 = 𝑏𝑢𝑑𝑔𝑒𝑡, 𝑝+ = 𝑝𝑟𝑖𝑐𝑒, 𝑟+= 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑟𝑒𝑡𝑢𝑟𝑛, 𝑥+ 𝜖 0, 1
• Weights: 𝜃=, 𝜃@, 𝜃A ≥ 0 s.t. 𝜃= + 𝜃@ + 𝜃A = 1
• ℎ+ = vector of historical price data
• 𝑐𝑜𝑣 ℎ+, ℎB =∑NOPe (RS,NTURS)(RV,NTRV)
fwhere 𝑛 = #of assets
𝒙𝒊 = 𝟏 → 𝒃𝒖𝒚 𝒙𝒊 = 𝟎 → 𝒅𝒐𝒏c𝒕 𝒃𝒖𝒚
Expected Return
6
Unconstrained Markowitz Formulation𝑚𝑎𝑥1
𝑓(𝑥)
𝑓 𝑥 = 𝜃=>+
𝑥+𝑟++𝑥+ − 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ − 𝜃A>+,B
𝑥+𝑐𝑜𝑣 ℎ+, ℎB 𝑥B
• 𝑏 = 𝑏𝑢𝑑𝑔𝑒𝑡, 𝑝+ = 𝑝𝑟𝑖𝑐𝑒, 𝑟+= 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑟𝑒𝑡𝑢𝑟𝑛, 𝑥+ 𝜖 0, 1
• Weights: 𝜃=, 𝜃@, 𝜃A ≥ 0 s.t. 𝜃= + 𝜃@ + 𝜃A = 1
• ℎ+ = vector of historical price data
• 𝑐𝑜𝑣 ℎ+, ℎB =∑NOPe (RS,NTURS)(RV,NTRV)
fwhere 𝑛 = #of assets
𝒙𝒊 = 𝟏 → 𝒃𝒖𝒚 𝒙𝒊 = 𝟎 → 𝒅𝒐𝒏c𝒕 𝒃𝒖𝒚
Budget Penalty
7
Unconstrained Markowitz Formulation𝑚𝑎𝑥1
𝑓(𝑥)
𝑓 𝑥 = 𝜃=>+
𝑥+𝑟++𝑥+ − 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ − 𝜃A>+,B
𝑥+𝑐𝑜𝑣 ℎ+, ℎB 𝑥B
• 𝑏 = 𝑏𝑢𝑑𝑔𝑒𝑡, 𝑝+ = 𝑝𝑟𝑖𝑐𝑒, 𝑟+= 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑟𝑒𝑡𝑢𝑟𝑛, 𝑥+ 𝜖 0, 1
• Weights: 𝜃=, 𝜃@, 𝜃A ≥ 0 s.t. 𝜃= + 𝜃@ + 𝜃A = 1
• ℎ+ = vector of historical price data
• 𝑐𝑜𝑣 ℎ+, ℎB =∑NOPe (RS,NTURS)(RV,NTRV)
fwhere 𝑛 = #of assets
𝒙𝒊 = 𝟏 → 𝒃𝒖𝒚 𝒙𝒊 = 𝟎 → 𝒅𝒐𝒏c𝒕 𝒃𝒖𝒚
Diversification
8
Fitting to Unconstrained Problem
• Assets can be divided into any desired fraction based on the budget.
• Normalize purchase price to the budget.
• Use Binary Fractional series: fractions=12q,12=,12@, … ,
12f Special thanks to Benjamin Stump
for the binary fractional series idea
9
Quantum Annealing for Unstructured Search
• Quantum annealing is a model of quantum computing tailored to unconstrained search
• QA operates by preparing a uniform superposition of all possible binary combinations.
• The initial state is then evolved toward a Hamiltonian that defines the objective function.
• Measurement samples the prepared probability distribution, which should concentrate at the global extrema.
10
Quantum Annealing for Unstructured Search
• D-Wave Systems provides a fourth-generation quantum annealer, 2000Q– Programmable superconducting integrated circuit– 2048-qubit register in a 2D Chimera layout– EM shielding, UHV, cooled to 14mK
• This is a special-purpose optimization solver that finds the energetic minimum of an Ising model– Search uses quantum annealing to find minima
• Non-zero temperature, flux noise, and other fluctuations complicate device physics
11
QUBO to Quantum Ising Hamiltonian
𝑓 𝑥 = −𝜃=>+
𝑥+𝑟++𝑥+ + 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ + 𝜃A>+,B
𝑥+𝑐𝑜𝑣 𝑝+, 𝑝B 𝑥B
𝑓 𝑥 = 𝑄+,B∼ 𝑥+𝑥B + 𝑞+𝑥+𝑞+ = 𝑄++ 𝑎𝑛𝑑 𝑄+,B∼ = 𝑄+,B 𝑖 ≠ 𝑗
Ising: 𝑦 𝜖 {−1, 1}
QUBO: 𝑥 𝜖 0, 1
𝐽+,B =14𝑄+,B∼ ℎ+ =
𝑞+2+>
B
𝐽+,B
𝛾 = =|∑+,B 𝑄+,B +
=@∑+ 𝑞+
}𝐻 =>+,B
𝐽+,B�̂�+�̂�B +>+
ℎ+ �𝑥+ + 𝛾
𝑚𝑖𝑛1𝑓(𝑥)
𝑓 𝑥 = >+
ℎ+𝑦+ +>+,B
𝐽+,B 𝑦+𝑦B + 𝛾
Coupler Strengths
12
QUBO to Quantum Ising Hamiltonian
𝑓 𝑥 = −𝜃=>+
𝑥+𝑟++𝑥+ + 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ + 𝜃A>+,B
𝑥+𝑐𝑜𝑣 𝑝+, 𝑝B 𝑥B
𝑓 𝑥 = 𝑄+,B∼ 𝑥+𝑥B + 𝑞+𝑥+𝑞+ = 𝑄++ 𝑎𝑛𝑑 𝑄+,B∼ = 𝑄+,B 𝑖 ≠ 𝑗
Ising: 𝑦 𝜖 {−1, 1}
QUBO: 𝑥 𝜖 0, 1
𝐽+,B =14𝑄+,B∼ ℎ+ =
𝑞+2+>
B
𝐽+,B
𝛾 = =|∑+,B 𝑄+,B +
=@∑+ 𝑞+
}𝐻 =>+,B
𝐽+,B�̂�+�̂�B +>+
ℎ+ �𝑥+ + 𝛾
𝑚𝑖𝑛1𝑓(𝑥)
𝑓 𝑥 = >+
ℎ+𝑦+ +>+,B
𝐽+,B 𝑦+𝑦B + 𝛾
Qubit Weights
13
QUBO to Quantum Ising Hamiltonian
𝑓 𝑥 = −𝜃=>+
𝑥+𝑟++𝑥+ + 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ + 𝜃A>+,B
𝑥+𝑐𝑜𝑣 𝑝+, 𝑝B 𝑥B
𝑓 𝑥 = 𝑄+,B∼ 𝑥+𝑥B + 𝑞+𝑥+𝑞+ = 𝑄++ 𝑎𝑛𝑑 𝑄+,B∼ = 𝑄+,B 𝑖 ≠ 𝑗
Ising: 𝑦 𝜖 {−1, 1}
QUBO: 𝑥 𝜖 0, 1
𝐽+,B =14𝑄+,B∼ ℎ+ =
𝑞+2+>
B
𝐽+,B
𝛾 = =|∑+,B 𝑄+,B +
=@∑+ 𝑞+
𝑚𝑖𝑛1𝑓(𝑥)
𝑓 𝑥 = >+
ℎ+𝑦+ +>+,B
𝐽+,B 𝑦+𝑦B + 𝛾
Constant }𝐻 =>+,B
𝐽+,B�̂�+�̂�B +>+
ℎ+ �𝑥+ + 𝛾
14
QUBO to Quantum Ising Hamiltonian
𝑓 𝑥 = −𝜃=>+
𝑥+𝑟++𝑥+ + 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ + 𝜃A>+,B
𝑥+𝑐𝑜𝑣 𝑝+, 𝑝B 𝑥B
𝑓 𝑥 = 𝑄+,B∼ 𝑥+𝑥B + 𝑞+𝑥+𝑞+ = 𝑄++ 𝑎𝑛𝑑 𝑄+,B∼ = 𝑄+,B 𝑖 ≠ 𝑗
Ising: 𝑦 𝜖 {−1, 1}
QUBO: 𝑥 𝜖 0, 1
𝐽+,B =14𝑄+,B∼ ℎ+ =
𝑞+2+>
B
𝐽+,B
𝛾 = =|∑+,B 𝑄+,B +
=@∑+ 𝑞+
𝑚𝑖𝑛1𝑓(𝑥)
𝑓 𝑥 = >+
ℎ+𝑦+ +>+,B
𝐽+,B 𝑦+𝑦B + 𝛾
Ising Form
}𝐻 =>+,B
𝐽+,B�̂�+�̂�B +>+
ℎ+ �𝑥+ + 𝛾
15
QUBO to Quantum Ising Hamiltonian
𝑓 𝑥 = −𝜃=>+
𝑥+𝑟++𝑥+ + 𝜃@>+
𝑥+𝑝+𝑥+ − 𝑏 @ + 𝜃A>+,B
𝑥+𝑐𝑜𝑣 𝑝+, 𝑝B 𝑥B
𝑓 𝑥 = 𝑄+,B∼ 𝑥+𝑥B + 𝑞+𝑥+𝑞+ = 𝑄++ 𝑎𝑛𝑑 𝑄+,B∼ = 𝑄+,B 𝑖 ≠ 𝑗
Ising: 𝑦 𝜖 {−1, 1}
QUBO: 𝑥 𝜖 0, 1
𝐽+,B =14𝑄+,B∼ ℎ+ =
𝑞+2+>
B
𝐽+,B
𝛾 = =|∑+,B 𝑄+,B +
=@∑+ 𝑞+
𝑚𝑖𝑛1𝑓(𝑥)
𝑓 𝑥 = >+
ℎ+𝑦+ +>+,B
𝐽+,B 𝑦+𝑦B + 𝛾
Quantum Ising
}𝐻 =>+,B
𝐽+,B�̂�+�̂�B +>+
ℎ+ �𝑥+ + 𝛾
16
Solving on D-Wave 2000QD-Wave 2000Q:• 2048 qubits
• Anneal times
• Embeddings
• Reverse Annealing
Parameters:
• 𝜃=, 𝜃@, 𝜃A• Number of assets
• Random historical Price Data
Outputs:• Portfolio:
[-1, 1, 1, 1, -1, -1…] à
[0, 1, 1, 1, 0, 0…]
• Energy of the portfolio
17
Benchmarking against Brute Force Solver
• Probability of success:
𝑃𝑂𝑆 = ���
• 𝑙 = # of ground state energies found by the D-Wave
• 𝑁� = number of samples/anneals
• Average Probability of success
𝑃𝑂𝑆 = ∑+�� ��� +
��• 𝑁� = # problems• 𝑖 indicates problem
• POS ratio: ��� ���� �
• 𝑃𝑂𝑆 � for reverse annealing• 𝑃𝑂𝑆 � for forward annealing
• Energy Binning
Energy à Energy Level
18
Best Fit for Probability of Success
𝑦 = 1.2245 𝑒Tq.��=|∗1�.����
• Embedding = Find_embedding()
• Slices = 0
• Anneal time = 𝟓𝝁𝒔
• Forward annealing
19
Forward Annealing Controls
20
Spin Reversal Control
21
Example Program Embedding in Hardware
find_embedding() Embedding
Clique Embedding
Problem size = 20
• 23 unit cells • 15 unit cells
22
Clique vs Find Embedding
Forward Annealing:
• Embeddings:• find_embedding()• Clique
• Anneal Time = 15𝜇𝑠• Spin Reversal = 0
23
Clique vs Find Embedding
Anneal time = 15 𝜇𝑠Number of anneals = 1,000
24
Forward Annealing: Clique vs Find Embedding
Anneal time = 15 𝜇𝑠Number of anneals = 1,000
25
Reverse Annealing
Parameters:
• Initial State• Reinitialize • S• Pause time
26
Reverse Annealing Sweep
27
Reverse Annealing Sweep
28
Reverse Annealing Sweep
29
Reverse Annealing Sweep
30
Overall Probability of Success
Reverse Annealing:
• S = .9• Pause = 15𝜇𝑠• I.S. = Forward• Reinitialize = False
Forward Annealing:
• Embeddings:• find_embedding()• Clique
• Anneal Time = 15𝜇𝑠• Spin Reversal = 0
31
Conclusions• Quantum annealing is a method for solving QUBO problems and
Markowitz Portfolio optimization is an example which demonstrates real-world application.
• Various parameters both in the QUBO and the D-Wave computer can be controlled/fine-tuned to yield better results.
• Forward annealing reveals a sub-exponential decrease in probability of success as problem size increases.
• Spin reversal transform improves variance.
• Clique Embedding improves probability of success over the find_embedding() method.
• Reverse annealing reveals a better probability of success and a better average solution if initial state is close to ground state.
32
Future Work
• Measure efficiency of using reverse annealing with forward annealing.
• Compare quantum annealing to classical heuristics.
• Investigate using real market data.
ORNL is managed by UT-Battelle, LLC for the US Department of Energy
Questions
University of Tennessee, Knoxville
Oak Ridge National Laboratory
Quantum Computing Institute
In collaboration with Khalifa University
Nada Elsokkary, Faisal Khan
34
Asseti
350
375
312
300
330
Asseti
1.0606
1.1363
.94545
.90909
1
Asseti, 0
1.0606
1.1363
.94545
.90909
1
Asseti, 1
.53030
.56815
.47273
.45450
.5
Asseti, 2
.26515
.28408
.23636
.22727
.25
𝒑𝒌𝒑𝒎 𝒂𝒊 → 𝒂𝒊,𝒌
35
Probability of Success
36
Optimality Gap
Optimality Gap:
𝑎𝑏𝑠𝐸¦§�𝐸¨1�
− 1
Recommended