Convex Optimization Over Risk-Neutral Probabilities
CConvex Optimization Over Risk-Neutral Probabilities
Shane Barratt Jonathan Tuck Stephen BoydMarch 9, 2020
Abstract
We consider a collection of derivatives that depend on the price of an underlyingasset at expiration or maturity. The absence of arbitrage is equivalent to the existenceof a risk-neutral probability distribution on the price; in particular, any risk neutraldistribution can be interpreted as a certificate establishing that no arbitrage exists. Weare interested in the case when there are multiple risk-neutral probabilities. We describea number of convex optimization problems over the convex set of risk neutral priceprobabilities. These include computation of bounds on the cumulative distribution,VaR, CVaR, and other quantities, over the set of risk-neutral probabilities. Afterdiscretizing the underlying price, these problems become finite dimensional convexor quasiconvex optimization problems, and therefore are tractable. We illustrate ourapproach using real options and futures pricing data for the S&P 500 index and Bitcoin.
The arbitrage theorem is a central result in finance originally proposed by Ross [31]. For amarket with a finite number of investments and possible outcomes, the arbitrage theoremstates that there either exists a probability distribution (called a risk-neutral probability ) overthe outcomes such that the expected return of all possible investments is nonpositive ( i.e. ,arbitrage does not exist), or there exists a nonnegative combination of the investments thatguarantees positive expected return ( i.e. , arbitrage exists). The no-arbitrage assumption isthat financial markets are arbitrage-free. For the most part, this holds, since if the marketswere not arbitrage-free, someone would take advantage of the arbitrage, changing the priceuntil it no longer exists. Under the no-arbitrage assumption, a notable implication of thearbitrage theorem is that a risk-neutral probability serves both as a conceivable distributionover the outcomes and as a certificate ensuring that arbitrage is impossible.For a given market, the set of risk-neutral probabilities is a polyhedron, and arbitrage isimpossible if the set is nonempty. We can verify that the market is arbitrage-free by findinga point in the feasible set of a particular system of linear equalities. Apart from verifyingthe no-arbitrage assumption, this set has many other uses. For example, it has been used forprojecting onto the set of risk-neutral probabilities using various distance measures ( e.g. , (cid:96) norm, (cid:96) norm, and KL-divergence) [32, 27, 33, 11, 24, 10], as well as for computing bounds1 a r X i v : . [ q -f i n . C P ] M a r n option prices given moments or other information [7, 6, 26]. These methods have beenapplied to various derivative markets, including equity indices [5, 3], currencies [12], andcommodities [28]. We consider nonparametric models of risk-neutral probabilities in thispaper; another viable option is to consider parametric models, i.e. , choose a distribution andfit its parameters to observed pricing data (see, e.g. , [ ? , 24] and the references therein). Wenote that once a risk-neutral distribution is found, it is often used to construct stochasticprocesses of the price of the underlying asset, e.g. , as a binomial tree [32, 23]. Risk-neutralprobabilities have also been used to infer properties of investor’s utility functions [4, 25].In this paper we consider the general problem of minimizing a convex or quasiconvexfunction over the (convex) set of risk-neutral probabilities. By considering convex optimiza-tion problems, finding a solution is tractable, and indeed has linear complexity in the numberof outcomes, which lets us scale the number of outcomes to the tens of thousands. More-over, the advent of domain-specific languages (DSLs) for convex optimization, e.g. , CVXPY[15, 2], make not just solving, but also formulating these problems straightforward; theyrequire just a few lines of a high-level language such as Python. We show that there aremany useful applications of convex optimization problems over risk-neutral probabilities,which encompass a lot of prior work, including computation of bounds on expected values ofarbitrary functions of the expiration price, estimation of the risk-neutral probability usingother information, computation of bounds on the cost of existing or new investments, andsensitivities of various quantities to the cost of each investment. We illustrate a number ofthese applications using real derivatives pricing data for the S&P 500 index and Bitcoin.There are a number of notable limitations to our approach. First, we require the num-ber of outcomes to be finite and reasonably small. Suppose, e.g. , that we tried to applyour approach to American-style options, which can be exercised at any time up until toexpiration. Even if we discretized the price of the underlying asset and time, the numberof outcomes would be exponentially large, since we would need to consider the price of theasset at each time point until expiration. (We note however that precise valuation and opti-mal exercise of American options is still mostly an unsolved problem.) Second, we considerstatic investments, i.e. , the investment is fixed until expiration. This precludes multi-periodinvestment models [16], dynamic hedging strategies that are at the core of derivative pricingmodels like the Black-Scholes model [8, 29], as well as treatment of American options, sincewe need to decide whether to exercise an option or not based on the current price. Despitethese limitations, we find that our approach can be very useful in practice and is also veryinterpretable, as demonstrated by our examples in § Outline.
The remainder of the paper is organized as follows. In § § §
4, we illustrateour approach on real derivatives pricing data for the S&P 500 Index and Bitcoin.2
Risk-neutral probabilities
Setting.
We consider a market for an asset, referred to as the underlying , with a numberof derivatives that provide payoffs at the same future date or time, referred to as expirationor maturity. We assume that there are n possible investments that include, e.g. , buying or(short) selling the underlying, as well as buying or selling (writing) derivatives. We let p > Payoff.
The payoff function f i : R + → R denotes the dollar amount received (or paid,if negative) per unit held of the i th investment; we give some examples of payoff functionsbelow. If we own a quantity w i ≥ i th investment, then at expiration, we wouldreceive f i ( p ) w i dollars. (We note that we do not discount payoffs at the risk-free rate, butwe could easily do this in our formulation.) Cost.
We let c ∈ R n denote the cost in dollars to acquire one unit of each investment( c i < w i ≥ i th investment, it would cost us c i w i dollars, and thereturn of our investment, at expiration, would be ( f i ( p ) − c i ) w i dollars. In this section we give some examples of payoff functions (see, e.g. , [22] for an overview ofvarious derivatives).
Underlying.
In many cases we can directly invest in the underlying. We allow both goinglong (buying), and going short (selling borrowed shares). Going long in the underlying hasa payoff function f ( p ) = p, and going short in the underlying has the payoff function f ( p ) = − p. European options.
A European option is a contract that gives one party the right tobuy or sell an underlying asset at an agreed upon strike price. If the right is to buy theunderlying asset (a call option), the option will only be exercised if the underlying price isgreater than the strike price. Conversely, if the right is to sell the underlying asset (a putoption), the option will only be exercised if the underlying price is less than the strike price.Under this logic, the payoff functions for buying European options with a strike price s are f call ( p ) = ( p − s ) + , f put ( p ) = ( p − s ) − , x + = max( x,
0) and x − = ( − x ) + . The payoff function for selling (writing) a Europeanoption is f w , call ( p ) = − f call ( p ) , f w , put ( p ) = − f put ( p ) . Futures.
Futures are contracts that obligate the buyer of the contract to buy or sell theunderlying asset at an agreed upon strike price. A long futures contract means the partymust buy the underlying asset at that strike price. Denoting the strike price of the futureby s , the payoff for buying a long futures contract is f ( p ) = p − s. A short futures contract means the party must sell the underlying asset at that strike price.The return function for buying a short futures contract is f ( p ) = s − p. Binary options.
A binary option is a contract that pays either a fixed monetary amountor nothing depending on the underlying’s price. For example, a binary option that pays thebuyer one dollar if the underlying asset is above a strike price s has a payoff function f ( p ) = (cid:40) p ≥ s, . If we sell that same binary option, the payoff function is f ( p ) = (cid:40) − p ≥ s, . For the remainder of the paper we will work with a discretized version of the price p , meaningit can only take one of m values p , . . . , p m , where we assume p < p < · · · < p m . (We notethat the discretization can be unequally spaced.) Since the methods that we describe involveconvex optimization, they scale well (and often linearly) with m [9]; this implies that m canbe chosen to be large enough that the discretization error is negligible.We can define a probability distribution over p as a vector π ∈ R m , with Prob ( p = p i ) = π i . Such a vector is in the set∆ = { π ∈ R m | π ≥ , T π = 1 } , i.e. , the probability simplex in R m . 4 .3 The set of risk-neutral probabilities Payoff matrix.
We can summarize the payoffs of each investment for each possible out-come with the payoff matrix P ∈ R m × n , with entries given by P ij = f j ( p i ) , i = 1 , . . . , m, j = 1 , . . . , n. Here P ij is the payoff in dollars per unit invested in investment j , if outcome i occurs. Arbitrage.
Let w ∈ R n + denote an investment vector, meaning we invest in a quantity w i of the i th investment, and hold these investments until expiration. The overall investmentwill cost us c T w now, and our expected payoff at expiration will be π T P w , meaning ourexpected return is ( P T π − c ) T w . Arbitrage is said to exist if there exists an investment vectorthat guarantees positive expected return, i.e. , there exists w ≥ P T π − c ) T w > P T π − c ) T w subject to w ≥ , (1)with variable w , is unbounded above. The set of risk-neutral probabilities.
We say that π is a risk-neutral probability (or no-arbitrage distribution ) if arbitrage is impossible, that is, if the optimal value of problem (1)is bounded. By LP duality [13] or the Farkas lemma [18], (1) is bounded if and only if P T π ≤ c . This means that the set of risk-neutral probabilities is the (convex) polyhedronΠ = { π ∈ ∆ | P T π ≤ c } . We note that if Π is empty, then arbitrage exists. We can interpret π ∈ Π as a distributionover the outcomes for which it is impossible to invest and receive positive expected return.
Another interpretation.
Consider the problemmaximize t subject to P w − ( c T w ) ≥ t ,w ≥ , (2)with variable w . (This problem is equivalent to problem (1).) If it is unbounded above, thenfor every R >
0, there exists an investment vector w that guarantees our return will be atleast R , no matter what the outcome is. The dual ismaximize 0subject to π ∈ Π , (3)with variable π . Therefore, another interpretation of π ∈ Π is as a certificate guaranteeingthat it is impossible to always have positive return regardless of the outcome.5 .
00 0 .
33 0 .
50 1 . π . . . . π simplex P T π ≤ P T π ≤ Figure 1: An example of the set of risk-neutral probabilities Π, denoted by the thick linesegment. Here P i denotes the i th column of P . Example.
Suppose there are n = 2 investments, m = 2 outcomes, the prices are c = (1 , P = (cid:20) / / / (cid:21) . Then the set of risk-neutral probability distributions isΠ = { ( x, − x ) | / ≤ x ≤ / } . We visualize the construction of this set in figure 1.
The general problem of convex optimization over risk-neutral probabilities isminimize L ( π )subject to π ∈ Π , (4)with variable π , where L : R m → R ∪ { + ∞} is convex (or quasiconvex). We use infinitevalues of L to encode constraints. Π is a polyhedron, so (4) is a convex optimization problem[9]. In general problem (4) does not have an analytical solution, but we can numerically findthe global optimum efficiently using modern convex optimization solvers [9]. All of theproblems we describe below (and many others) are readily expressed in a few lines of codeusing domain specific languages for convex optimization, such as CVX [20, 21], CVXPY[15, 2], Convex.jl [34], or CVXR [19]. 6 .1 Functions of the price Suppose g : R → R is some function of the underlying’s price at expiration; the expectationof g is E g ( p ) = m (cid:88) i =1 π i g ( p i ) , which is a linear function of π . Some examples of functions of the price include: • The price.
Here g ( p ) = p . The expected value is the expected price. • The return on an investment.
Here g ( p ) = (cid:80) ni =1 ( f i ( p ) − c i ) w i for an investment w ∈ R n + . The expected value is the expected return of the investment. • Indicator functions of arbitrary sets.
Here g ( p ) = 1 if p ∈ C and 0 otherwise, for someset C ⊆ R . The expected value is Prob ( p ∈ C ). Bounds on expected values.
We can compute lower and upper bounds on expectedvalues of functions of the price by respectively letting L ( π ) = E g ( p ) and L ( π ) = − E g ( p )and solving problem (4). For example, we could compute bounds on the expected price orthe return on a given investment. Bounds on ratios of expected values.
If we have another function f of the price, and g ( p ) >
0, then the function E f ( p ) E g ( p ) = (cid:80) i π i f ( p i ) (cid:80) i π i g ( p i ) , is quasilinear. We can find bounds on this ratio by minimizing and maximizing this quantity,both of which are both quasiconvex optimization problems. For example, we can computebounds on Prob ( p ∈ A | p ∈ B ) for two sets A ⊆ R and B ⊆ R , since it is equal to Prob ( p ∈ A ∩ B ) Prob ( p ∈ B ) . CDF.
The cumulative distribution function (CDF) of g is the function F ( x ) = Prob ( g ( p ) ≤ x ) = (cid:88) g ( p i ) ≤ x π i , which, for each x , is linear in π . For example, if g ( p ) = p , then F ( x ) is just the CDF ofthe price. We can compute lower and upper bounds on the CDF at x = p , . . . , p m , byminimizing and maximizing F ( x ) subject to π ∈ Π.7 aR.
The value-at-risk of g ( p ) at probability (cid:15) ∈ [0 ,
1] is defined as
VaR ( g ( p ); (cid:15) ) = inf { α | Prob ( g ( p ) ≤ α ) ≥ (cid:15) } = F − ( (cid:15) ) , where F − ( (cid:15) ) = inf { x | F ( x ) ≥ (cid:15) } [17]. From the bounds on the CDF, we can computebounds on the value at risk as F − ( (cid:15) ) ≤ VaR ( g ( p ); (cid:15) ) ≤ F − ( (cid:15) ) . CVaR.
The conditional value-at-risk of g ( p ) at probability (cid:15) is defined as (see, e.g. , [30]) CVaR ( g ( p ); (cid:15) ) = inf β (cid:18) β + E ( g ( p ) − β ) + − (cid:15) (cid:19) = min i (cid:32) p i + m (cid:88) j =1 π j ( g ( p j ) − p i ) + − (cid:15) (cid:33) , which is a concave function of π . Therefore, we can find an upper bound on CVaR by letting L ( π ) = − CVaR ( g ( p ); (cid:15) ). Since conditional value-at-risk is bounded below by value-at-risk, F − ( (cid:15) ) is a (trivial) lower bound. Constraints.
We can incorporate upper or lower bounds on the expected values of func-tions of the price as linear equality constraints in the function L . These linear inequalityconstraints can be interpreted as adding another investment. For example, if we add theconstraint a T π ≤ b for a ∈ R m and b ∈ R , this is the same as if we had originally includedan investment with a payoff function f ( p i ) = a i and cost b . Maximum entropy.
We can find the maximum entropy risk-neutral probability by letting L ( π ) = m (cid:88) i =1 π i log( π i ) . Minimum KL-divergence.
Given another distribution η ∈ ∆, we can find the closestrisk-neutral probability distribution to η as measured by Kullback-Leibler (KL) divergenceby letting L ( π ) = m (cid:88) i =1 π i log( π i /η i ) . Closest log-normal distribution.
We can approximately find the closest log-normaldistribution to Π by performing the following alternating projection procedure, startingwith π ∈ Π: • Fit a log-normal distribution to π k with mean and variance µ = m (cid:88) i =1 π i log( p i ) , σ = m (cid:88) i =1 π i (log( p i ) − µ ) . Discretize this distribution, resulting in η k ∈ ∆. • Set π k +1 equal to the closest risk-neutral probability distribution to η k , in terms ofKL-divergence. If π k +1 is close enough to η k , then quit.For better performance, this process may be repeated for various π ∈ Π. Suppose that we want to add another investment, and would like to come up with lowerand upper bounds on its cost subject to the constraint that arbitrage is impossible, i.e. ,there exists a risk-neutral probability distribution. Suppose the payoff function of the newinvestment is f ( p i ) = ( p new ) i , where p new ∈ R m . We can find lower and upper bounds onthe cost of this new investment by respectively letting L ( π ) = p T new π and L ( π ) = − p T new π and solving problem (4). Bertsimas and Popescu [6, §
3] were among the first to proposecomputing bounds on option prices based on prices of other options.
Validation.
We can check whether our prediction is accurate by holding out each invest-ment one at a time and comparing the lower and upper bounds that we find with the trueprice.
Suppose L is convex and let λ (cid:63) ∈ R n + denote the optimal dual variable for the constraint P T π ≤ c in problem (4), and let L (cid:63) ( c ) denote the optimal value as a function of c . A global inequality.
For ∆ c ∈ R n , the following global inequality holds [9, § L (cid:63) ( c + ∆ c ) ≥ L (cid:63) ( c ) − ( λ (cid:63) ) T ∆ c. Local sensitivity.
Suppose L (cid:63) ( c ) is differentiable at c . Then ∇ L (cid:63) ( c ) = − λ (cid:63) [9, § c ∈ R n will decrease L (cid:63) by roughly( λ (cid:63) ) T ∆ c . We implemented all examples using CVXPY [15, 2], and each required just a few lines ofcode. The code and data for all of these examples have been made freely available online at .1 Standard & Poor’s 500 Index In our first example, we consider the Standard & Poor’s 500 index (SPX) as the underlying,which is a market-capitalization-weighted index of 500 of the largest publicly traded U.S.companies, and excludes dividends. We gathered the end-of-day (EOD) best bid and askprice of all SPX options on June 3, 2019, as well as the price of the index, which was 2744.45dollars, from the OptionMetrics Ivy database via the Wharton Research Data Services [1].We discretized the expiration price from 1500 to 4000 dollars, in 50 cent increments,resulting in m = 5000 outcomes. We allowed six possible investments: buying or sellingputs, buying or selling calls, and buying or selling the underlying. The payoffs for each ofthese investments are described in §
2. The cost of each investment is the ask price if buying,the negative bid price if selling, plus a 65 cent fee for buying/selling each option (which atthe time of writing are the fees for the TD Ameritrade brokerage), and a 0 .
3% fee for buyingor selling the underlying.We consider the options that expire 25 days into the future, on June 28, 2019. Therewere 112 puts and 81 calls expiring on June 28 that had non-empty order books, i.e. , hadat least one bid and ask quote. Therefore, we allow n = 2(112 + 81) + 2 = 388 investments. Functions of the price.
We calculated bounds on the expected value of the expirationprice. The lower bound was 2745.77 dollars and the upper bound was 2747.03 dollars. Wethen computed bounds on the probability that the expiration price is 20% below the currentprice, given that the expiration price is less than the current price; this probability was foundto be between 0 .
4% and 2%. We also computed bounds on the CDF, complementary CDF(CCDF), and VaR of the expiration price, and plot these figure 5 we plot these bounds.
Estimation.
We computed the maximum entropy risk-neutral distribution, as well as the(approximately) closest log-normal distribution to the set of risk-neutral probabilities. Theclosest log-normal distribution was log( p ) ∼ N (7 . , . Bounds on costs.
We held out each put and call option one at a time and computedbounds on their bid and ask prices. In figure 4 we plot our computed lower and upperbounds along with the true prices. We observe that the bounds seem to be quite tight, andindeed bound the observed prices.
In our next example, we consider the crypto-currency Bitcoin as the underlying. As deriva-tives, we use Deribit European-style options and futures, whose underlying is the Deribit10
000 2200 2400 2600 2800 3000 3200 3400 x . . . . . . P r o b ( p ≤ x ) upperlowercurrent price (a) Bounds on Prob ( p ≤ x ). . . . . . . (cid:15) p upperlower (b) Bounds on VaR ( p, (cid:15) ). x . . . . . . P r o b ( p ≥ x ) upperlowercurrent price (c) Bounds on Prob ( p ≥ x ). Figure 2: SPX example. Bounds on CDF,
VaR , and CCDF.11
000 2200 2400 2600 2800 3000 3200 3400 p . . . . p . . . . Figure 3: SPX example. Left: maximum entropy risk-neutral distribution; Right: closestlog-normal distribution. pu t a s k p r i ce lower boundtrue (a) Put ask prices. pu t b i dp r i ce upper boundtrue (b) Put bid prices. − c a ll a s k p r i ce lower boundtrue (c) Call ask prices. − c a ll b i dp r i ce upper boundtrue (d) Call bid prices. Figure 4: SPX example. Bounds on costs.12TC index, which is the average of six leading BTC-USD exchange prices: Bitstamp, Bit-trex, Coinbase Pro, Gemini, Itbit, and Kraken. We gathered the prices of March 27, 2020Bitcoin options and futures on February 20, 2020 using the Deribit API [14].We discretized the expiration price from 5 to 30000 dollars, in 5 dollar increments, re-sulting in m = 6000 outcomes. We allow six possible investments: buying or selling puts,buying or selling calls, and buying or selling futures. The cost of each investment is the askprice if buying, the negative bid price if selling, plus a 0.04% fee for option transactions, a0.075% fee for buying futures, and a 0.025% (market-maker) rebate for selling futures (whichat the time of writing are the fees for the Deribit exchange).In total, there were 16 puts and 19 calls expiring on March 27, with strike prices rangingfrom 4000 to 18000. This means there were n = 2(16 + 19) + 2 = 72 possible investments. Functions of the price.
We calculated bounds on the expected value of the expirationprice. The lower bound was 9847.7 dollars and the upper bound was 9852.57 dollars. Wealso computed bounds on the CDF, CCDF, and the value-at-risk of the expiration price. Infigure 5 we plot these bounds.
Estimation.
We computed the maximum entropy risk-neutral distribution, as well as the(approximately) closest log-normal distribution to the set of risk-neutral probabilities. Theclosest log-normal distribution was log( p ) ∼ N (9 . , . . Bounds on costs.
We held out each put and call option one at a time and computedbounds on their bid and ask prices. In figure 7 we plot our computed lower and upperbounds along with the true prices. We observe that the bounds are quite tight, and indeedbound the observed prices.
Sensitivities.
We computed the optimal dual variable of the constraint P T π ≤ c for theentropy maximization problem. In table 1 we list the five largest dual variables, along withtheir corresponding investments and costs. We observe that shorting the underlying, as wellas buying/writing various calls have the most effect on the maximum entropy risk-neutralprobability. For example, if we decrease the price of the 9000 call by ten dollars, then theentropy will decrease by at least 0.01. In this paper we described applications of minimizing a convex or quasiconvex function overthe set of convex risk-neutral probabilities. These include computation of bounds on thecumulative distribution, VaR, conditional probabilities, and prices of new derivatives, as13
000 6000 8000 10000 12000 14000 16000 x . . . . . . P r o b ( p ≤ x ) upperlowercurrent price (a) Bounds on Prob ( p ≤ x ). . . . . . (cid:15) p upperlower (b) Bounds on VaR ( p, (cid:15) ). x . . . . . . P r o b ( p ≥ x ) upperlowercurrent price (c) Bounds on Prob ( p ≥ x ). Figure 5: Bitcoin example. Bounds on CDF,
VaR , and CCDF.14 p . . . . . . . p . . . . . . Figure 6: Bitcoin example. Left: maximum entropy risk-neutral distribution; Right: closestlog-normal distribution. pu t a s k p r i ce lower boundtrue (a) Put ask prices. pu t b i dp r i ce upper boundtrue (b) Put bid prices. c a ll a s k p r i ce lower boundtrue (c) Call ask prices. c a ll b i dp r i ce upper boundtrue (d) Call bid prices. Figure 7: Bitcoin example. Bounds on costs.15able 1: Bitcoin example. Dual variables for entropy maximization problem.Investment c i λ (cid:63)i Short underlying -9847.65 0.001Buy 9000 call 1191.268 0.001Write 18000 call -19.69 0.0004Buy 10000 call 659.63 0.0004Buy 8000 call 1973.96 0.0002well as estimation problems. We reiterate that all of the aforementioned problems can betractably solved, and due to DSLs, are easy to implement. A potential avenue for futureresearch is use the set of risk-neutral probabilities for multiple expiration dates to somehowconnect the distribution of the underlying’s price movements between those dates.16 cknowledgements
Data from Wharton Research Data Services was used in preparing this paper. S. Barratt issupported by the National Science Foundation Graduate Research Fellowship under GrantNo. DGE-1656518. J. Tuck is supported by the Stanford Graduate Fellowship in Science &Engineering.
References [1] Wharton Research Data Services: OptionMetrics. .[2] A. Agrawal, R. Verschueren, S. Diamond, and S. Boyd. A rewriting system for convexoptimization problems.
Journal of Control and Decision , 5(1):42–60, 2018.[3] J. ¨Aij¨o. Impact of US and UK macroeconomic news announcements on the returndistribution implied by FTSE-100 index options.
International Review of FinancialAnalysis , 17(2):242–258, 2008.[4] Y. Aıt-Sahalia and A. Lo. Nonparametric risk management and implied risk aversion.
Journal of Econometrics , 94(1-2):9–51, 2000.[5] D. Bates. Post-’87 crash fears in the S&P 500 futures option market.
Journal ofEconometrics , 94(1-2):181–238, 2000.[6] D. Bertsimas and I. Popescu. On the relation between option and stock prices: A convexoptimization approach.
Operations Research , 50(2):358–374, 2002.[7] D. Bertsimas and J. Tsitsiklis.
Introduction to linear optimization . Athena Scientific,1997.[8] F. Black and M. Scholes. The pricing of options and corporate liabilities.
Journal ofPolitical Economy , 81(3):637–654, 1973.[9] S. Boyd and L. Vandenberghe.
Convex Optimization . Cambridge University Press, 2004.[10] N. Branger. Pricing derivative securities using cross-entropy: An economic analysis.
International Journal of Theoretical and Applied Finance , 7(01):63–81, 2004.[11] P. Buchen and M. Kelly. The maximum entropy distribution of an asset inferred fromoption prices.
The Journal of Financial and Quantitative Analysis , 31(1):143–159, 1996.[12] O. Castr´en. Do options-implied RND functions on G3 currencies move around the timesof interventions on the JPY/USD exchange rate? 2004.[13] G. Dantzig.
Linear programming and extensions . Princeton University Press, 1998.1714] Deribit. Deribit API V2.0.0. https://docs.deribit.com/v2/ .[15] S. Diamond and S. Boyd. CVXPY: A Python-embedded modeling language for convexoptimization.
Journal of Machine Learning Research , 17(83):1–5, 2016.[16] D. Duffie.
Dynamic asset pricing theory . Princeton University Press, 2010.[17] D. Duffie and J. Pan. An overview of value at risk.
The Journal of Derivatives , 4(3):7–49,1997.[18] J. Farkas. Theorie der einfachen ungleichungen.
Journal fur die reine und angewandteMathematik , 124:1–27, 1902.[19] A. Fu, B. Narasimhan, and S. Boyd. CVXR: An R package for disciplined convexoptimization. In
Journal of Statistical Software , 2019.[20] M. Grant and S. Boyd. Graph implementations for nonsmooth convex programs. In
Recent Advances in Learning and Control , Lecture Notes in Control and InformationSciences, pages 95–110. Springer-Verlag Limited, 2008.[21] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming,version 2.1, 2014.[22] J. Hull.
Options, futures, and other derivatives . Pearson Prentice Hall, 2006.[23] J. Jackwerth. Generalized binomial trees.
Journal of Derivatives , 5(2):7–17, 1996.[24] J. Jackwerth. Option-implied risk-neutral distributions and implied binomial trees: Aliterature review.
The Journal of Derivatives , 7(2):66–82, 1999.[25] J. Jackwerth. Recovering risk aversion from option prices and realized returns.
TheReview of Financial Studies , 13(2):433–451, 2000.[26] J. Jackwerth. Option-implied risk-neutral distributions and risk aversion. 2004.[27] J. Jackwerth and M. Rubinstein. Recovering probability distributions from optionprices.
The Journal of Finance , 51(5):1611–1631, 1996.[28] W. Melick and C. Thomas. Recovering an asset’s implied PDF from option prices: anapplication to crude oil during the gulf crisis.
Journal of Financial and QuantitativeAnalysis , 32(1):91–115, 1997.[29] R. Merton. Theory of rational option pricing.
The Bell Journal of Economics andManagement Science , pages 141–183, 1973.[30] T. Rockafellar and S. Uryasev. Optimization of conditional value-at-risk.
Journal ofRisk , 2:21–42, 2000. 1831] S. Ross.
Return, risk and arbitrage . Rodney L. White Center for Financial Research,The Wharton School, 1973.[32] M. Rubenstein. Implied binomial trees.
The Journal of Finance , 49(3):771–818, 1994.[33] M. Stutzer. A simple nonparametric approach to derivative security valuation.
TheJournal of Finance , 51(5):1633–1652, 1996.[34] M. Udell, K. Mohan, D. Zeng, J. Hong, S. Diamond, and S. Boyd. Convex optimizationin Julia.