## Crude oil prices density forecast on July, 5, 2005 for end-September 2005.

## Source details

**Canonical URL:** [Crude oil prices density forecast on July, 5, 2005 for end-September 2005.](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2005/_wp05174.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2005/_wp05174.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2005/_wp05174.pdf.json)

---

### Introduction and motivation
- Treat oil as an asset in addition to a commodity; derivatives markets (futures, options) influence oil price behavior and can exert pressure on prices and raise volatility.
- Financial investors (institutional investors, hedge funds, commercial entities) generate sizable demand for futures contracts and speculative activity that can:
  - Exert pressure on prices.
  - Cause volatility of oil futures prices to rise to excessive levels.
- High volatility stimulates speculation, which contributes to higher volatility and volatility clustering.
- Policymakers and central bankers monitor options prices to gauge market sentiment and forecast price distributions because options reflect both consumption and investment aspects of oil as an asset.
- Purpose: model asset-price uncertainty and option pricing using Lévy processes that handle discontinuities and behave properly under time aggregation.

### Key features motivating Lévy-process modeling
- Lévy processes allow:
  - Jumps at random times (small jumps more frequent than large jumps).
  - Asymmetries and fat tails in asset returns.
  - Infinitely divisible distributions that can be skewed and have slowly decaying tails.
- Empirical phenomena addressed:
  - Implied volatility smiles/smears and stochastic volatility.
  - Non-normality of returns (stable laws, subordinated processes).
- Paper structure (overview):
  - Section II: Lévy processes and Levy-Khintchine formula.
  - Section III: subordination methodology and examples.
  - Section IV: market incompleteness and Esscher transform.
  - Section V: pricing using characteristic functions.
  - Section VI: application to crude oil options; density forecasts show high volatility and right skew.
  - Section VII: conclusion.

### Lévy processes — definitions, properties, and characteristic functions
- Lévy process Xt (t ≥ 0) is cadlag with X0 = 0 and:
  - independent increments;
  - stationary increments;
  - stochastic continuity.
- Infinitely divisible: for any t > 0, integer n ≥ 2, Xt = sum_{k=1}^{n} (X_{t k/n} − X_{t (k−1)/n}).
- Levy-Khintchine formula for cumulant characteristic function ψ(u):
  - ψ(u) = iγu − (1/2)σ^2 u^2 + ∫_{R\{0}} (e^{i u x} − 1 − i u x 1_{|x|<1}) ν(dx),
    where γ ∈ R, σ^2 ≥ 0, and ν is a Lévy measure with ∫_{−∞}^{∞} (1 ∧ x^2) ν(dx) < ∞.
- Lévy-Ito decomposition: Xt = γ t + σ W_t + J_t + M_t (Brownian component plus jumps driven by a Poisson point process).
- Finite-activity vs. infinite-activity:
  - Finite-activity: ∫_{R\{0}} ν(dx) < ∞ (e.g., Merton compound Poisson).
  - Infinite-activity: ∫_{R\{0}} ν(dx) = ∞ (e.g., NIG, VG, CGMY).

### Subordination and time-changed Lévy processes; examples
- Time change: X_t = Z_{T_t} where Z_t is a Lévy process and T_t is an increasing stochastic clock (subordinator).
  - Produces stochastic volatility and can capture leverage and activity effects.
  - Clark (1973): subordinated Brownian motion by trading volume/number of transactions yields finite-variance, semi-heavy-tailed distributions.
  - Monroe (1978): every semimartingale can be written as a time-changed Brownian motion (equivalently time-changed Lévy process).
- Characteristic function of subordinated process:
  - If Z_t has characteristic exponent ψ(u) and T_t has Laplace exponent L(·), then X_t = Z_{T_t} has CF E[e^{i u X_t}] = exp(t L(ψ(u))).
- Parametric classes discussed:
  - Inverse Gaussian (IG) and Generalized Inverse Gaussian (GIG) subordinators (CFs and Lévy measures involving modified Bessel functions).
  - Variance-Gamma (VG): VG_t = θ G_t + σ W_{G_t} where G_t is Gamma process with parameters 1/υ and 1/(υ); VG CF φ_{VG}(u; t, σ, υ, θ) = [1 − i θ u υ + (1/2) σ^2 u^2 υ]^{−t/υ}; VG has infinite activity.
  - Normal Inverse Gaussian (NIG): Brownian motion subordinated by IG; semi-heavy tails with asymptotic decay f_{NIG}(x; α, β, δ) ~ exp(−α |x| + β x) |x|^{−3/2} as x→±∞ (up to constant).
  - Generalized Hyperbolic (GH): normal variance-mean mixture with GIG mixing distribution; includes VG and NIG as special cases.

### Market incompleteness and Esscher transform
- Lévy models are typically incomplete: infinite number of equivalent martingale measures (EMMs).
- Esscher transform (parameter h) defines an equivalent measure by tilting densities:
  - f(x;t;h) = e^{h x} f(x;t) / M(h;t), where M is the MGF.
  - Choose h* so discounted price e^{−r t} S_t is a martingale under Q_{h*}: M(1+h*; 1) = e^{r} M(h*; 1).
  - CF under Esscher transform: φ_h(u) = φ(u − i h) / φ(−i h); infinite divisibility preserved.
- Esscher measure linked to minimum entropy martingale measure in cited literature.

### Option pricing using characteristic functions and FFT
- Option pricing via state-price density and characteristic functions:
  - CFs are often available in closed form even when densities are not, enabling Fourier-space pricing and numerical inversion.
  - Lewis (2001) formulation: option value as a complex integral using conditional CF φ_{z}(τ).
  - Carr and Madan (1999) FFT approach:
    - Use damped option price c̃(k) = e^{α k} C(k) (α > 0) so Fourier transform exists.
    - ψ_T(v) = e^{−r T} φ_{log(S_T)}(v − (α + 1) i) / (α^2 + α − v^2 + i v (2 α + 1)).
    - Option price recovered by inverse Fourier transform and numerically approximated by N-point discrete FFT with spacing Δ; choose α to control truncation.

### Application to crude oil options — calibration and inverse problem
- Calibration problem: infer risk-neutral distribution from observed option prices by estimating parameter vector Φ of exponential Lévy model S_t = S_0 e^{X_t} under risk-neutral measure.
- Calibration by minimizing quadratic pricing error:
  - Φ̂ = arg min_Φ (1/N) ∑_{i=1}^{N} (Ĉ(K_i, T; Φ) − C_i)^2,
    subject to put-call parity: Put_t + S_t = Call_t + K e^{−r (T − t)}.
- Empirical calibration example:
  - Data: observed crude oil options on July, 5, 2005 for maturity end-September 2005.
  - Risk-free rate used: three-month U.S. treasury bill rate.
  - Crude futures price reported as 59.81 per barrel.
  - Estimation yielded parameter estimates: 2 ˆ σ = 0.32, ˆ υ = 0.20, ˆ θ = 0.10.

### Estimation results, implied density, and robustness caveats
- Reported parameter interpretations:
  - 2 ˆ σ = 0.32 — indicates high volatility characterizing the oil market.
  - ˆ υ = 0.20 — shows fat tails, implying higher probability than the normal distribution for important deviations from the futures level.
  - ˆ θ = 0.10 — indicates positive skewness; market assigns higher probability to upward deviations from the expected mean.
- Robustness and calibration caveats:
  - Robustness checked using linear model A·D·q = , where A is vector of call and put prices, D is pay-off matrix, q is vector of Arrow–Debreu prices.
  - Because of high volatility, market expectations can change dramatically intra-day or from one day to another, changing the density forecast for a given maturity.
  - Calibration results should be interpreted with caution.

### Market structure, behavior, and drivers (empirical notes)
- Composition of large traders:
  - Aggregate large-traders’ positions reported to the CFTC usually represent 70-90 percent of total open interests in a derivatives market.
  - Data for February 1, 2005: commercial traders held 67.1 percent of the open long positions and 69.2 percent of the short positions in crude oil futures on NYMEX.
- Behavioral dynamics:
  - Participants include hedgers, arbitrageurs, speculators, institutional investors (pension funds), hedge funds, and commercial entities registered with the CFTC.
  - Commercial traders sometimes take speculative short-term positions during large price swings.
  - Very low interest rates reduce the cost of shorting bonds and margin requirements, increasing futures market volume and potentially exerting upward pressure on futures prices.
  - A significant portion of crude oil price increases could be attributed to derivatives markets and to the role of crude oil as an asset.

### Modeling implications and methodological advantages
- Lévy-process-based models are well suited to:
  - Capture skewness, fat tails, and stochastic volatility observed in high-frequency financial data.
  - Model frequent and large jumps caused by supply bottlenecks or demand shocks.
  - Use subordination (business time) to relate returns to market activity/news rather than calendar time.
- Tractability: closed-form characteristic functions enable efficient option pricing (Fourier inversion, FFT) and recovery of probability distributions from CFs.
- Market incompleteness necessitates selection of a martingale measure (e.g., Esscher transform) for pricing.

### Policy-relevant findings and recommendations
- Calibration/extraction findings:
  - Risk-neutral distribution extracted from options indicated positive skew: higher probability mass on crude oil prices remaining above the futures’ level — consistent with sustained upward pressure on oil prices.
  - Lévy market models provide useful tools for the Fund’s work: analyzing high-frequency data, gauging market sentiment, and designing policy responses.
- Suggested policy assessments and actions:
  - Assess factors causing pressure on crude oil demand, including low interest rates and depreciating currencies.
  - Seek greater energy efficiency and inter-energy substitution.
  - Monitor speculative activity in derivatives markets; high speculative activity associated with high volatility leads to volatility clustering and greater uncertainty in futures prices.
  - Incorporate both the asset and commodity roles of crude oil in energy modeling; Lévy processes and inverse problems provide a framework to assess both aspects.

### Appendix: Empirical Characteristic Function (ECF) estimation — summary
- Motivation:
  - Lévy-type models often lack tractable densities and can have unbounded likelihoods; the CF is bounded, may have closed form, and retains full sample information.
- ECF/GMM approach:
  - Match model CF and empirical CF over a grid {u_j} in the Fourier domain; q grid points yield 2q moment conditions (real and imaginary parts).
  - Typically 2q > l where l is the parameter vector dimension.
  - Under regularity and sufficiently large/fine grids, the ECF estimator can be made arbitrarily asymptotically efficient (Feuerverger and McDunnough (1981a)).
- Asymptotic properties and implementation:
  - ECF estimator ˆθ is strongly consistent and asymptotically normal with covariance involving V, Ω, and derivatives of V with respect to θ.
  - ECF minimization: find θ minimizing (Vn − V(.;θ))′ ˆΩ−1 (Vn − V(.;θ)); equivalent to second-stage GMM and GMM-efficient.
  - With an optimal weight (Fourier transform of the score), ECF is equivalent to maximum likelihood; optimal weight is typically unknown when likelihood has no closed form.
- Practical guidance:
  - Asymptotic efficiency depends critically on choice of grid points {u_j}; sufficiently fine and extended grids may achieve the Cramer-Rao lower bound in some cases.

*Source: _wp05174 (PDF chapter/section).*

### 1.  Crude oil prices density forecast on July, 5, 2005 for end-September 2005. .....................20

### _wp05174 - 1.  Crude oil prices density forecast on July, 5, 2005 for end-September 2005

### Crude oil prices density forecast (July, 5, 2005 for end-September 2005)
- Title of content unit: "Crude oil prices density forecast on July, 5, 2005 for end-September 2005."
- Located on page 20 of the PDF.
- Content focus implied by title: density forecast for crude oil prices targeting end-September 2005 based on information as of July, 5, 2005.
- No numerical forecasts or probability density values are present in the supplied content excerpt.

### Appendix: Empirical Characteristic Function and Estimation in the Frequency Domain
- Appendix title: "Empirical Characteristic Function and Estimation in the Frequency Domain."
- Located starting on page 22.
- Implied methodological focus: use of the empirical characteristic function and frequency-domain estimation techniques relevant to the forecast or model.

### References
- References section indicated on page 24.
- No reference entries are present in the supplied excerpt.

### Document structure and pagination notes
- Table of contents entries shown:
  - "1.  Crude oil prices density forecast on July, 5, 2005 for end-September 2005. .....................20"
  - "Appendix Empirical Characteristic Function and Estimation in the Frequency Domain...............22"
  - "References................................................................................................................................24"
- Internal marker: "- 3 -"
- Section marker: "I."

*Source: _wp05174 - 1.  Crude oil prices density forecast on July, 5, 2005 for end-September 2005. (PDF), canonical URL provided by the source metadata.*

### INTRODUCTION

### _wp05174 - INTRODUCTION

### Introduction and motivation
- Treat oil as an asset in addition to a commodity; recognize influence of derivatives markets on oil price behavior.
- Financial investors (institutional investors, hedge funds, commercial entities) generate sizable demand for futures contracts and speculative activity that can:
  - Exert pressure on prices.
  - Cause volatility of oil futures prices to rise to excessive levels.
- High volatility in commodity and currency markets stimulates speculation, which in turn contributes to higher volatility and to volatility clustering.
- Policymakers and central bankers monitor derivatives markets, particularly options prices, to gauge market sentiment and forecast price distributions because these markets account for both consumption and investment aspects of an asset.
- Purpose of paper: model asset prices’ uncertainty and option pricing in the context of Levy processes, capable of handling discontinuities and behaving properly under time aggregation.

### Key features motivating Levy-process modeling
- Levy processes overcome shortcomings of the Black-Scholes (1973) diffusion framework by allowing:
  - Jumps at random times (small jumps more frequent than large jumps).
  - Asymmetries and fat tails in asset returns.
  - Infinitely divisible probability distributions that can be skewed and have slowly decaying tails.
- Empirical observations:
  - Implied volatility varies with state and time (volatility smile or smirk); volatility may be stochastic.
  - Mandelbrot (1963) and Fama (1965) noted non-normality of returns; stable distributions were proposed but have infinite variance.
  - Clark (1973) proposed subordinated processes (time-changed Brownian motion) to obtain finite-variance, semi-heavy-tailed distributions (e.g., using volume or number of transactions as subordinator).

### Structure and objectives of the paper
- Section II: describe Levy processes and their characteristic function (Levy-Khintchine formula).
- Section III: methodology for constructing Levy processes through subordination; examples of distributions obtained through subordination that fit financial time series.
- Section IV: market incompleteness under Levy jumps and the Esscher transform for selecting a martingale measure.
- Section V: methodology for pricing contingent claims using characteristic functions.
- Section VI: application to crude oil options to infer density forecasts of future oil prices; findings point to high volatility of oil futures prices and a right-skewed market expectation (greater probability mass on upward deviations from the mean).
- Section VII: conclusion.

### Levy processes — definitions and properties
- Levy process (LP) definition: cadlag (right continuous and left limit, RCLL) stochastic process {Xt, t ≥ 0} with X0 = 0 and:
  - (i) independent increments;
  - (ii) stationary increments;
  - (iii) stochastic continuity: for every ε > 0, lim_{h→0} P(|X_{t+h} − X_t| ≥ ε) = 0.
- LPs are limits of random walks and are infinitely divisible into i.i.d. random variables:
  - For any t > 0 and integer n ≥ 2, Xt = sum_{k=1}^{n} X_{t k/n} − X_{t (k−1)/n}.
- Common infinitely divisible laws: Gaussian, Gamma, α−stable, Poisson.

### Levy-Khintchine formula and characteristic exponent
- If φ(u) is the characteristic function (CF) of Xt, the cumulant characteristic function ψ(u) = log φ(u) satisfies the Levy-Khintchine formula:
  - ψ(u) = iγu − (1/2)σ^2 u^2 + ∫_{R\{0}} (e^{i u x} − 1 − i u x 1_{|x|<1}) ν(dx),
    where γ ∈ R is drift, σ^2 ≥ 0 is volatility, and ν is a Levy measure on R\{0} with ∫_{−∞}^{∞} (1 ∧ x^2) ν(dx) < ∞.
- The Levy density f(x) (if ν(dx) = f(x) dx) describes local uncertainty of a pure jump process analogous to instantaneous volatility for diffusions.
- Levy-Ito decomposition: Xt = γ t + σ W_t + J_t + M_t, decomposing LP into Brownian motion with drift and jump components; jumps follow a Poisson point process with intensity ν.

### Finite-activity vs. infinite-activity jumps
- Finite-activity pure-jump LP: ∫_{R\{0}} ν(dx) < ∞ (aggregate jump arrival rate finite); classical example: Merton (1976) compound Poisson with jump magnitude N(α, σ^2).
  - Merton Levy measure: ν(dx) = λ (1 / (σ √(2π))) exp( − (x − α)^2 / (2 σ^2) ) dx.
- Kou (2002): double-exponential conditional jump size; Levy measure given accordingly.
- Infinite-activity processes: ∫_{R\{0}} ν(dx) = ∞; examples: NIG (Barndorff-Nielsen, 1998), generalized hyperbolic (Eberlein et al., 1998), variance-gamma (VG) (Madan and Milne, 1991), CGMY (Carr et al., 2002).

### Subordination and time-changed Levy processes
- Time change: run an LP on a stochastic, increasing clock T_t (subordinator) to capture jumps, stochastic volatility, and leverage effect.
- Clark (1973): subordinated processes (X_{T_t})—using volume or number of transactions as T_t—produce finite-variance, semi-heavy-tailed distributions; kurtosis of increments of X_{T_t} increases with variance of increments of T_t.
- Monroe (1978): every semimartingale can be written as a time-changed Brownian motion; equivalently, every semimartingale can be written as a time-changed LP.
- Time-changing a Brownian motion with drift μ by an increasing process T_t yields Xt = μ T_t + σ W_{T_t}, producing stochastic volatility when observed in calendar time.

### Characteristic function of subordinated processes
- If Z_t is an LP with characteristic exponent ψ(u) and T_t an independent subordinator with Laplace exponent L(u), then X_t = Z_{T_t} has CF:
  - E[e^{i u X_t}] = E[e^{T_t ψ(u)}] = exp(t L(ψ(u))) — CF is the Laplace transform of T evaluated at ψ(u).
- Introducing correlation between innovations in Z and the time change T can introduce asymmetry (skewness).

### Examples of subordinators and time-changed processes (parametric classes)
- Inverse Gaussian (IG) process:
  - IG(a,b) CF: exp(a b (1 − √(1 − 2 i u / b^2)) ) in the form provided.
  - Density and Levy measure expressed; IG is infinitely divisible and defines an IG process with stationary independent increments.
- Generalized Inverse Gaussian (GIG) process:
  - GIG(a,b,λ) density and CF involve modified Bessel function K_λ.
  - Levy measure density expressed via integrals involving Bessel functions.
- Variance-Gamma (VG) process:
  - VG_t = θ G_t + σ W_{G_t}, where G_t is a Gamma process with parameters 1/υ and 1/(υ).
  - VG CF: φ_{VG}(u; t, σ, υ, θ) = [1 − i θ u υ + (1/2) σ^2 u^2 υ]^{−t/υ} (as expressed).
  - VG moments: mean = θ; variance = σ^2 υ + θ^2; skewness = (3 θ υ σ + 2 θ^3) / (σ^2 υ + θ^2)^{3/2} (presented in text as formula); kurtosis given by explicit formula in text.
  - VG has infinite-activity (Levy measure has infinite mass).
- Normal Inverse Gaussian (NIG) process:
  - NIG obtained by subordinating Brownian motion with an IG process; CF and Levy measure expressed with modified Bessel functions.
  - NIG has infinite activity and semi-heavy tails: f_{NIG}(x; α, β, δ) ~ exp(−α |x| + β x) |x|^{−3/2} as x→±∞ up to multiplicative constant.
- Generalized Hyperbolic (GH) process:
  - GH is normal variance-mean mixture with mixing distribution GIG; CF and density given in terms of Bessel functions.
  - GH includes VG and NIG as special/reparameterized cases; hyperbolic (HYP) when υ = 1; NIG when υ = −1/2.

### Implications and empirical relevance of time-changed Levy models
- Time-changed pure-jump infinite-activity models (NIG, VG, HYP/GH) empirically capture skewness, leptokurtosis, and implied volatility smiles better than classical diffusion or jump-diffusion models.
- Tractability of characteristic functions enables option pricing via Fourier inversion and fast numerical methods (FFT).
- Probability distributions can be recovered from CF via numerical inversion:
  - F(x) = 1/2 + (1/π) ∫_{0}^{∞} Im(e^{−i u x} φ(u) / u) du (or related inversion formula as presented).

### Market incompleteness and Esscher transform
- Except when X is Brownian motion or a Poisson process, Levy models are incomplete: infinite number of equivalent martingale measures (EMMs) compatible with absence of arbitrage.
- Esscher transform (Gerber and Shiu, 1994):
  - For LP Xt with density f(x;t) and MGF M(u;t), the Esscher density with parameter h is f(x;t;h) = e^{h x} f(x;t) / M(h;t).
  - Choose h* so that the discounted price process e^{−r t} S_t is a martingale under the Esscher-transformed measure Q_{h*}.
  - Martingale condition: M(1+h*; 1) = e^{r} M(h*; 1) (equivalently log relationships provided).
  - Under Esscher change, CF transforms as φ_h(u) = φ(u − i h) / φ(−i h), preserving infinite divisibility; characteristic triplet transforms accordingly (explicit transformations provided).
- Esscher transform can be identified with the minimum entropy martingale measure (Miyahara, 2004).
- Alternative approach (Carr et al., 2003): choose u such that E[e^{i u X_t}] is a martingale for appropriate u; risk-neutral CF of log price constructed from expectation expressions.

### Option pricing using characteristic functions
- Option value under martingale pricing is convolution of discounted payoff with state price density; state price density relates to second derivative of call prices (Breeden and Litzenberger, 1978).
- For many Levy processes, transition densities are complicated; CF often available in closed form, enabling Fourier-space computation of option prices.
- Lewis (2001) formulation: option value expressed as a complex integral (convolution in Fourier space) using conditional CF of state price density φ_{z}(τ).
- Carr and Madan (1999) FFT approach:
  - Use damped option price c̃(k) = e^{α k} C(k) (α > 0) so Fourier transform exists.
  - Relationship: ψ_T(v) = e^{−r T} φ_{log(S_T)}(v − (α + 1) i) / (α^2 + α − v^2 + i v (2 α + 1)) (as structured in text).
  - Option price recovered by inverse Fourier transform and approximated numerically via N-point discrete FFT with grid spacing Δ; Simpson's or trapezoidal integration weights used.
- Parameter α chosen to control truncation error and ensure boundedness.

### Application to crude oil options — inverse problem and calibration
- Inverse problem: infer risk-neutral density (model calibration) from observed option prices by backing out parameters Φ of exponential Levy model S_t = S_0 e^{X_t} under risk-neutral measure with triplet ((·), (·), (·))_{σ ν γ}.
- Calibration via minimizing quadratic pricing error:
  - Φ̂ = arg min_Φ (1/N) ∑_{i=1}^{N} (Ĉ(K_i, T; Φ) − C_i)^2,
    subject to put-call parity: Put_t + S_t = Call_t + K e^{−r (T − t)}.
  - C_Φ denotes model call price for exponential Levy model with parameters Φ; C_i denotes observed call prices for maturity T and strikes K_i.
- Empirical calibration example (text):
  - Calibration applied to the Variance-Gamma model.
  - Data set: observed crude oil options on July 5, 2005 for maturity end-September 2005.
  - Risk-free interest: three-month U.S. treasury bill rate.
  - Crude futures price: US$ (value not provided in supplied excerpt).
- Findings summary (from Section VI overview):
  - Extracted densities from options point to high volatility of oil futures prices and a right-skewed market expectation (greater probability mass on upward deviations from the mean).
  - Extracting densities from options allows analysis of crude oil's role as an asset as well as a commodity and improves oil market modeling.

*Source: _wp05174 - INTRODUCTION (PDF chapter/section) — content provided from the canonical IMF source document*

### 59.81 per barrel. The estimation yielded the following parameters:

### _wp05174 - 59.81 per barrel. The estimation yielded the following parameters:

### Estimation results and implied density
- Reported parameter estimates: 2 ˆ σ = 0.32, ˆ υ = 0.20, ˆ θ = 0.10.
- Interpretation of parameters:
  - 2 ˆ σ indicates high volatility characterizing the oil market.
  - ˆ υ shows fat tails, implying higher probability than the normal distribution for important deviations of prices from the futures level.
  - ˆ θ indicates positive skewness, meaning the market was according higher probability for upward deviations from the expected mean.
- Caveats on robustness and calibration:
  - Robustness confirmed by use of a linear model A·D·q = , where A is a vector of call and put options prices, D is a pay-off matrix, and q is a vector of Arrow-Debreu prices.
  - Owing to high volatility, market expectations can change dramatically intra-day or from one day to the other and thus can change dramatically the density forecast for a given maturity time.
  - Calibration results need to be interpreted with caution.

### Market structure, behavior, and drivers
- Participant composition and positions:
  - The aggregate of all large-traders’ positions reported to the CFTC usually represents 70-90 percent of total open interests in any given derivatives market.
  - Data for February 1, 2005 indicated that commercial traders held 67.1 percent of the open long positions and 69.2 percent of the short positions in crude oil futures on the New York Mercantile Exchange (NYMEX).
- Behavioral dynamics and market effects:
  - Traders in derivatives markets are hedgers, arbitrageurs, and speculators; many types of investors participate, including institutional investors (e.g., pension funds) and hedge funds and commercial entities registered with the CFTC.
  - Commercial traders occasionally take speculative short-term positions during periods of large price swings.
  - High volatility and volatility clustering increase speculative activity and add pressure on futures prices.
  - Very low interest rates reduce the cost of shorting bonds and the cost of margin requirements and increase volume in the futures market; increased demand for long contracts would exert upward pressure on futures prices.
  - A significant portion of crude oil price increases could be attributed to derivatives markets and to the role of crude oil as an asset rather than as a commodity.

### Modeling implications and methodology
- Lévy processes and commodity vs. asset roles:
  - The use of Lévy processes and their corresponding inverse problem would allow study of the role of asset markets in the behavior of crude oil prices.
  - Commodity-side factors remain important: when demand acts against a short-term fixed crude oil supply and bottlenecks in refining and distribution capacity, frequent and large jumps in crude oil prices occur, suggesting the use of Lévy processes for modeling these jumps.
- Advantages of Lévy-process-based option pricing:
  - Lévy processes offer better tools for analyzing skewness, fat tails, and stochastic volatility in high-frequency financial data than classical diffusions or jump-diffusion models.
  - The concept of subordination measures returns relative to activity/news (business time) rather than calendar time; high activity or important news may cause higher volatility in returns.
  - Normal Inverse Gaussian, Variance-Gamma, and General Hyperbolic motions are Lévy processes that are time-changed Brownian motions, pure jump, and infinite activity models; their empirical performance in modeling skewness, leptokurtosis, and implied volatility smiles was deemed consistent with data.
  - Lévy processes lead to incomplete markets and an infinite number of martingale measures compatible with no arbitrage; the Esscher measure is one procedure to obtain a martingale measure.
  - Characteristic functions (CFs) and Fourier transforms are efficient tools for option pricing when CFs are available in closed form.

### Policy-relevant findings and recommendations
- Calibration findings and policy relevance:
  - The inverse-problem extraction of a risk-neutral distribution from crude oil options indicated market expectations were positively skewed—higher probability mass on crude oil prices remaining above the futures’ level—consistent with sustained upward pressure on oil prices.
  - The Lévy market model is highly relevant for the Fund’s work: it provides tools for analyzing high-frequency data, gauging market sentiment, and designing policy responses.
- Suggested policy assessments and actions:
  - Assess factors causing pressure on crude oil demand, including low interest rates and depreciating currencies.
  - Seek greater energy efficiency and inter-energy substitution.
  - Recognize the importance of derivatives markets in influencing crude oil prices; monitor speculative activity since high speculative activity associated with high volatility leads to volatility clustering and greater uncertainty in futures prices.
  - Energy modeling should incorporate both the role of crude oil as an asset and as a commodity; Lévy processes and inverse problems provide a framework to assess both aspects.

### Appendix summary — Empirical Characteristic Function (ECF) estimation
- Motivation:
  - Lack of tractable probability density functions and unbounded likelihood functions for Lévy-type models make maximum likelihood estimation difficult.
  - The characteristic function (CF), a Fourier transform of the density, is always bounded, can have closed-form expressions, and retains full sample information.
- ECF/GMM approach:
  - The ECF procedure matches the model CF and the empirical CF (ECF) obtained from data over a grid of Fourier-domain points; equivalent in essence to Generalized Method of Moments (GMM).
  - Using q discrete grid points yields 2q moment conditions (real and imaginary parts); typically 2q > l, where l is the dimension of parameter vector θ.
  - Under regularity conditions and sufficiently large samples and fine/extended grids, the ECF estimator can be made arbitrarily asymptotically efficient (Feuerverger and McDunnough (1981a)).
- Asymptotic properties and implementation:
  - The ECF estimator ˆθ is strongly consistent and asymptotically normal with an asymptotic covariance matrix given in the text (involving V, Ω, and derivatives of V with respect to θ).
  - The ECF minimization problem is to find θ that minimizes (Vn − V(.;θ))′ ˆΩ−1 (Vn − V(.;θ)), where ˆΩ is a consistent estimator of Ω; this is equivalent to a second-stage GMM estimator and yields GMM-efficient estimators.
  - The ECF procedure is also equivalent to maximum likelihood when an optimal weight (the Fourier transform of the score) is used; however, when the likelihood has no closed form, the optimal weight is unknown.
- Practical guidance:
  - The asymptotic efficiency of the ECF depends essentially on the choice of grid points {uj}; selecting the grid sufficiently fine and extended can, in some cases, achieve the Cramer-Rao lower bound for efficiency.

*Source: _wp05174 (PDF chapter/section).*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2005/_wp05174.pdf_