## 1. Time-Series Properties of the Oil Price

## Source details

**Canonical URL:** [1. Time-Series Properties of the Oil Price](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06299.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06299.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06299.pdf.json)

---

### I. Introduction — context and objectives
- Sample period analyzed: January 2, 2002–July 7, 2006 (1130 daily observations).
- Focus: dynamics of daily oil futures prices (Brent, three-month delivery; source: Reuters).
- Key motivations:
  - Pricing derivatives and hedging.
  - Policy making and short-term forecasting.
  - Lack of extensive modeling research on crude oil prices compared to equity indices.

### Major empirical observations (high-level)
- Price run-up:
  - Oil price rose from US$21.13/bl on January, 2002 to US$73.76/bl on July, 7, 2006.
  - Record level reached: US$78.30 per barrel (bl) on August 7, 2006.
- Price path characteristics:
  - Ascent was not monotonic or smooth: repeated peaks, temporary retreats, resumed higher peaks.
  - Movements dominated by high-intensity jumps; oil markets appeared constantly out-of-equilibrium.
- Volatility and market expectations:
  - Implied volatility in the range of 30 percent.
  - Market implied distributions were right-skewed (higher probabilities for further price increases).
  - Options activity indicated expectations of large upward jumps (price and volume of options at strikes in the range of US$75–US$85/bl).

### II. Empirical aspects of futures oil prices during 2002–06
- Time-series properties:
  - Augmented Dickey-Fuller test on level price (Null: St has a unit root):
    - Augmented Dickey-Fuller test statistic = -0.52; probability value = 0.88.
    - Test critical values: 1% level (-3.44); 5% level (-2.86); 10% level (-2.57).
  - Augmented Dickey-Fuller test on first differences (Null: ΔSt has a unit root):
    - Augmented Dickey-Fuller test statistic = -35.98; probability value = 0.00.
    - Test critical values: 1% level (-3.44); 5% level (-2.86); 10% level (-2.57).
  - AR2 specification reported (as presented in source):
    - S_t = 1.12 S_{t-1} - 0.31 S_{t-2} + 0.98? (source notation).
    - Coefficients shown in source: 12 31.32.30.98 0.930.070.09 (presented in table form).
    - R^2 = 0.99, DW = 2.05.
  - Stationarity: levels possess a unit root; changes ΔS_t are stationary.
- Macroeconomic/market fundamentals during sample:
  - World crude oil supply: rigid at about 84 million barrels per day (mbd) for most of sample.
  - World economy growth: about 4–5 percent in real terms during 2002–2006.
  - Demand price elasticity range: -0.01 to -0.04 (reference to IMF WP/06/62).
  - Short-run income elasticity estimates: between 0.2 and 0.4.
  - Implication: price inelastic demand and rigid supply imply small excess demand requires large price changes to clear markets.

### Distributional properties of daily log-price returns
- Return definition: x_t = log S_t - log S_{t-1}.
- Empirical distribution features:
  - Mode around 1–2 percent.
  - Daily changes in the range of 5–7 percent were not uncommon.
  - Frequency of jumps exceeding ±3 percent estimated at 23 percent.
  - Descriptive statistics:
    - mean = 0.116
    - standard deviation = 2.29 (annualized to 36.3 percent)
    - skewness = -0.39
    - kurtosis = 4.79
    - Jarque-Bera normality statistic = 179.3; probability-value = 0.0
  - Interpretation:
    - Left-skewed distribution: downward jumps of smaller size more frequent than upward jumps of larger size, but positive mean implies larger upward jumps outweigh smaller downward ones.
    - Fat tails: large jumps occur more frequently than under normality.
  - Conclusion: presence of skewness and kurtosis suggests modeling via jump-diffusion or general Lévy processes.

### B. Oil price time-varying volatility
- Two volatility measures analyzed:
  - Implied volatility from Brent futures options (August 2005–June 2006; source: Bloomberg).
    - Average implied volatility ~30 percent.
    - Volatility often surged to 34–35 percent.
  - GARCH(1,1) estimated volatility (January 2, 2002–July 7, 2006).
    - Return definition used: x_t = log S_t - log S_{t-1}.
    - GARCH mean volatility estimated at 43 percent.
    - Observed features: high volatility, volatility clustering, volatility rising during large price shocks and receding during price retreats.
- Market implications:
  - High volatility indicates large uncertainty and sensitivity to news and shocks.
  - Volatility clustering increases speculative demand for futures, which can further raise volatility.

### III. Crude oil price as a Merton jump-diffusion (J-D) process — model structure
- Continuous-time SDE for spot price S_t (J-D specification given in source):
  - dS_t = S_t [α dt + σ dB_t + (e^{J_t} - 1) dN_t]
  - Notation and assumptions:
    - α = instantaneous return; σ^2 = instantaneous variance.
    - B_t = standard Brownian motion with dB_t ~ (0, dt).
    - N_t = Poisson counter with intensity λ; arrivals independent of Brownian motion.
    - Jump size J_t iid ~ N(β, δ^2).
    - When a jump occurs, S_t = S_{t-} e^{J_t}; percentage change = e^{J_t} - 1.
- Log-price dynamics (using X_t = log S_t and Ito’s lemma):
  - dX_t = (α - 1/2 σ^2) dt + σ dB_t + ∑_{i=1}^{N_t} J_i
  - Drift parameter μ defined in source: μ = α - 1/2 σ^2.
  - Parameter vector: θ = (μ, σ^2, λ, β, δ^2).
- Discrete-time formulation (Δ = 1 day in presentation):
  - x_t = X_t - X_{t-1} = μ + diffusion component + sum of jumps over interval.
  - Density of x_t is a convolution of Gaussian diffusion and compound Poisson jump components:
    - Probability density (Δ = 1) expressed as an infinite sum over n = 0,1,2,... with Poisson weights and normal densities (as presented in source).
- Motivations for jumps:
  - Presence of skewness and kurtosis in returns.
  - Short-term implied volatility skews across strikes in option markets.
  - References cited: Merton (1976); Bakshi et al. (1997); Bates (1996).

### IV. Alternative estimation approaches (mentioned)
- Methods identified for estimating the J-D model:
  - Maximum Likelihood.
  - Method of Cumulants.
  - Method of Characteristic Function.
- Details on estimation procedures and parameter estimates appear in subsequent sections of the source document.

---

### 1. The maximum likelihood method: Let

### Maximum likelihood (ML) method — application issues and history
- Log-likelihood for an observed sample of T log returns xt, t = 1,...,T is given (in the source) as a double-sum expression involving μ, β, λ, σ, δ.
- Application issues:
  - Difficulties in identification of the jump parameter and instability of parameter estimates.
  - Computational tractability, un-boundedness over the parameter space, and instability of parameters have led researchers to alternative methods (method of moments, cumulants, CF-based methods).
- Historical applications:
  - Ball and Torous (1983) applied ML truncating the number of jumps at n = 15.
  - Ball and Torous (1985) and Jorion (1988) applied ML assuming a Bernoulli process for the jump component.
  - ML estimates achieve the lower bound for the Cramer-Rao efficiency criterion but practical problems motivate alternatives.

### Method of cumulants — key formulas and historical restrictions
- Characteristic function (CF) of Xt: φX(u) = E[exp(iuXt)] = ∫ exp(iuX) fX(X) dX.
- Cumulants κn are coefficients in power series expansion of ln φX(u).
- CF for the J-D process yields first four cumulants:
  - κ1 = μ + λβ
  - κ2 = σ2 + λδ2 + λβ2
  - κ3 = λβ3 + 3λβδ2
  - κ4 = λ(β4 + 6β2δ2 + 3δ4)
- Press (1967) imposed μ = 0 and derived closed-form relations for β̂, λ̂, δ̂, σ̂ in terms of sample cumulants; Press’ estimates were often wrong-signed and not plausible.
- Beckers (1981) set β = 0 and used sixth-order cumulants to derive estimators; improved over Press but anomalies remained.
- Ball and Torous (1983) used a Bernoulli jump process and Beckers’ restriction β = 0 to derive a system of cumulant equations and closed-form expressions for μ̂, λ̂, σ̂2, δ̂2 (explicit forms given in the source).

### Method of the characteristic function (CF) and empirical CF (ECF)
- CF and the corresponding probability density contain equivalent information; CF is often available in closed form even when transition densities are not.
- Method of moments/GMM and ECF procedures:
  - Moments of any order n can be computed from derivatives of the CF: En[Xn] = i−n (d nφ(u)/du n)|u=0.
  - ECF: empirical φn(u) = (1/T) ∑_{j=1}^T exp(iu xj) = (1/T) ∑ (cos(uxj) + i sin(uxj)).
  - Estimator θ̂ = argminφn − φθ' W (φn − φθ), where W is positive semi-definite.
  - ECF is essentially equivalent to GMM; with sufficiently large sample and fine grid in u, ECF estimators can reach arbitrarily high asymptotic efficiency (Feuerverger 1990).
  - Over-identification (r > l) allows tests of over-identifying restrictions (Hansen (1982) type tests).

### Empirical estimation results (Brent futures daily sample) — common findings
- Methods applied: ML (Bernoulli jump process), cumulants (various restrictions), ECF (unrestricted and with λ = 0.23).
- Common empirical findings across methods:
  - Dynamics influenced by both diffusion and jump components; jump component is dominant.
  - Jump component has much higher variance than the diffusion component, consistent with excess kurtosis.
  - Mean of jump size tends to be negative, consistent with negative skewness of returns.
  - Diffusion drift is high, consistent with observed upward trend in crude oil prices.
  - Jump dynamics: high intensity and large variance indicate frequent large discontinuities; markets often out-of-equilibrium.

### Selected parameter estimates (as reported in the source text; numeric forms preserved)
- Maximum Likelihood (Bernoulli jump process, unrestricted or with λ restricted in variants):
  - μ̂ = 0.23 (paragraph text)
  - 2σ̂ = 4.46
  - 2δ̂ = 4.47
  - λ̂ = 0.59
  - β̂ = −1.12
- Method of cumulants (with λ = 0.23):
  - μ̂ = 0.32
  - 2σ̂ = 1.81
  - 2δ̂ = 2.78
  - β̂ = −0.85
- Press (1967) restriction μ = 0 produced an implausible estimate:
  - 2δ̂ = −13.88
- Beckers (1981) restriction β = 0:
  - μ̂ = 0.12
  - 2σ̂ = 3.34
  - 2δ̂ = 8.62
  - λ̂ = 0.22
- ECF (unrestricted):
  - μ̂ = 0.57
  - 2σ̂ = 0.54
  - 2δ̂ = 0.96
  - λ̂ = 4.37
  - β̂ = −0.10
- ECF with λ = 0.23:
  - μ̂ = 0.27
  - 2σ̂ = 3.34
  - 2δ̂ = 6.97
  - β̂ = −52

- Interpretation provided in the source:
  - Estimates concordant: oil price process dominated by jump process with large discontinuities at high intensity.
  - Negative mean of jump component interpreted as short-lived downward adjustments in demand after large upward jumps in price.
  - Large positive diffusion drift reflects a strong income effect that pushes oil prices upward, dominating the negative mean of jump component for daily data.
  - Economic explanation: World demand highly elastic to world income (which expanded 4–5 percent per year during the sample), highly inelastic to price; supply rigid — small excess demand leads to large price variation; income effect dominates price effect.

---

### Variance-Gamma (VG) Lévy process — motivation, definition, and estimation

### Motivation for VG over J-D
- Limitations of J-D model:
  - Does not capture time-varying (stochastic) volatility.
  - Models finite large jumps but cannot capture infinite small jumps analogous to diffusion.
- VG Lévy process proposed to capture stochastic volatility and small frequent jumps while retaining tractable CF.

### VG definition and characteristic function (as in source)
- VG process: XVG t = α Gt + σ B_{Gt}, where {Gt, t ≥ 0} is a gamma process (mean rate 1/υ and variance rate 1/υ), B is Brownian motion, σ > 0, α ∈ R.
- CF of VG: φVG t(u; σ, υ, α) = [1 − iαuυ + (1/2)σ2u2υ]^{-t/υ} (equivalently as displayed in the source).
- VG parameters and moments (under (σ, υ, α)) — source expressions preserved:
  - Mean = α
  - Variance = σ2υ + α2
  - Skewness = (source expression preserved: skewness = (2 α υ + 3α(σ2υ + α2)?) as in source)
  - Kurtosis = (source expression preserved: kurtosis = 4222 3(1   2()   )υυσσ υα−+− + as displayed)
- Role of parameters:
  - α controls skewness (0α< implies negative skewness).
  - υ controls kurtosis and frequency of jumps (larger υ → more frequent jumps, fatter tails).

### 2. Estimation of the Variance-Gamma process — model & ECF estimates
- Price model: St = S0 exp(μt + Xt), where Xt is a Variance-Gamma (VG) process.
- Log return: Δxt = log(St) − log(St−1) = μ + Xt − Xt−1.
- Characteristic function (CF) of the log price return preserved from source notation.
- Estimation method: Empirical Characteristic Function (ECF) approach (see Table 3 in source).
- Textual summary of ECF findings:
  - Estimated parameters of the VG process were stable and statistically significant, corroborating the J-D process results.
  - Crude oil prices exhibited a high drift coefficient, high volatility, frequent and large jumps, and skewness.
  - The drift coefficient asserted a strong upward pulling force that kept reigniting oil prices.
  - Volatility parameter was high and significant, indicating high uncertainties about future price movements.
  - Parameter υ (tail fatness) was high and significant, implying frequent and large jumps.
  - Parameter α (skewness) was negative and significant, showing left-skewness in the VG distribution; interpretation: large positive jumps were typically followed by a cooling-off period with small/frequent negative jumps, while higher income demand elasticity (vs. price elasticity) meant faster world economic growth pushed demand upward.
- Explicit parameter values reported in paragraph text (preserve exact appearance):
  - ˆ μ = 0.09
  - ˆ σ = 1.66
  - ˆ υ = 0.80
  - ˆ α = −0.36  (textual line: "α, which controls for skewness, estimated at ˆ 0.36 α=−, was negative and significant")
- Table 3 parameter values (as presented in the table):
  - Drift μ: 0.09 (t = 3.52)
  - Drift α: −0.146 (t = −2.54)
  - Volatility σ: 1.66 (t = 80.4)
  - Variance of VG υ: 0.804 (t = 52.3)
  - Note: table footnote — "1/ Using ECF method."

### Option pricing using characteristic functions (methodology and formulas)
- Option pricing in asset price space: option value f(t,S t) = EQt[e−∫tT r(s) ds g(ST) | St = S] (source presents the convolution of discounted payoff with state price density).
- CF definition for log-price under risk-neutral measure: φTt(u) = ∫ p(ST | St) eiu log(ST) dST.
- Two Fourier-space pricing approaches preserved from source:
  - Heston (1993): numerical inversion of the CF (singularity at u = 0 problematic).
  - Carr and Madan (1999): numerical inversion of the Fourier transform of the option value (damped option price approach).
- Carr-Madan transform notation (preserved):
  - Damped option price: cT(k) ≡ exp(a k) CT(k) for a > 0.
  - Fourier transform of damped option price: ψT(u) = ∫ eiu k cT(k) dk.
  - Relation between ψT(u) and φT(u) (as given):  
    22 ((1)) () (21) rT T T euai u aauia u φ ψ − −+ = +− + + .
- FFT inversion and numerical quadrature approximations for option prices preserved in source notation.

### Risk-neutral calibration (inverse problem) and density forecast for crude oil
- Inverse problem: back out risk-neutral VG parameters θ = (σ, α, υ) by minimizing quadratic pricing error over N observed option prices, subject to put-call parity constraint.
- Objective (preserved formula): θ̂ = arg min(θ) (1/N) Σj=1..N [CTK* (·,θ) − CTK(·)]^2 under put-call parity: SP + CTK* − CTK = Ke−rT.
- Regularized minimization with penalty parameter h > 0 (preserved formula):  
  θ̂ = arg min (1/N) Σj [CTK* − CTK]^2 + h [e−rT S0 + CTK − PTK − Ke−rT]^2.
- Data and calibration specifics (preserve exact figures and dates):
  - Data date: July 21, 2006.
  - Options: call and put futures options contracts maturing end-September 2006.
  - Risk-free interest rate used (three-month U.S. Treasury bill rate): 4.965.
  - Crude futures price: US$74.43/bl.
- Constrained minimization output (risk-neutral triplet reported):
  - ˆ σ = 1.72
  - ˆ υ = 1.12
  - ˆ α = 0.37
- Interpretation of implied risk-neutral distribution (July 21, 2006 for end-September 2006 maturity):
  - Market participants expected no short-term change in underlying fundamentals.
  - Expected high volatility: ˆ σ = 1.72.
  - Dominated by a jump process: ˆ υ = 1.12.
  - Right-skewed risk-neutral distribution (ˆ α = 0.37), indicating higher probability mass for prices rising above the futures price than falling below it.

---

### Main conclusions, implications, and annexed relations

### Main conclusions and implications (sample: January 2, 2002 – July 7, 2006)
- Core empirical findings:
  - Dynamics dominated by frequent jumps causing markets to be often out-of-equilibrium.
  - Strong positive drift persisted, continually pushing prices upward despite retreats after major upward jumps.
  - High volatility made oil prices very sensitive to small shocks and news.
  - Both J-D and VG specifications produced consistent findings aligned with oil-market fundamentals and world economic conditions.
  - Faster world economic growth and expansionary monetary policies expanded crude demand; given price inelasticities of demand and supply, small excess demand/supply led to large price adjustments — explaining high jump intensity and upward pressure.
- Policy and practical relevance:
  - Results are relevant for forecasting, risk management, derivatives pricing, and gauging market sentiment.
  - Findings help establish the stochastic-process nature of oil prices and link process parameters to underlying fundamentals — implying parameter changes require fundamental shifts.
  - Relevance for the Fund in monitoring oil markets and seeking policies to stabilize these markets.

### Annex summary: Method of cumulants (key relations preserved)
- Moment generating function M(u) = E[e uX ] = ∫ fX(x) e u x dx.
- Cumulant-generating function: log M(u) = Σn=0..∞ κn u n / n!.
- Relation between raw moments mj = E[X^j] and cumulants κn via recursion (preserved form):  
  1 1 1 1 n nn n n j j n mm j κ κ − − − = − ⎛⎞ =− ⎜⎟ − ⎝⎠ Σ.
- Selected relations preserved:
  - m1 = κ1
  - m2 = κ2 + κ1^2
  - Central moment relations: first central moment 0, second central moment = κ2, third central moment = κ3, fourth central moment = κ4 + 3 κ2^2.
- Kurtosis measure preserved: kurtX = E[X^4] − 3 (E[X^2])^2 (source formula expressed).

*Source: _wp06299 - 1–2. Time-Series Properties of the Oil Price; The maximum likelihood method; Estimation of the Variance-Gamma process (IMF working paper chapter text provided).*

### 1. Time-Series Properties of the Oil Price .............................................................................

### 1. Time-Series Properties of the Oil Price

### I. Introduction — context and objectives
- Sample period analyzed: January 2, 2002–July 7, 2006 (1130 daily observations).
- Focus: dynamics of daily oil futures prices (Brent, three-month delivery; source: Reuters).
- Key motivations:
  - Pricing derivatives and hedging.
  - Policy making and short-term forecasting.
  - Lack of extensive modeling research on crude oil prices compared to equity indices.

### Major empirical observations (high-level)
- Price run-up:
  - Oil price rose from US$21.13/bl on January, 2002 to US$73.76/bl on July, 7, 2006.
  - Record level reached: US$78.30 per barrel (bl) on August 7, 2006.
- Price path characteristics:
  - Ascent was not monotonic or smooth: repeated peaks, temporary retreats, resumed higher peaks.
  - Movements dominated by high-intensity jumps; oil markets appeared constantly out-of-equilibrium.
- Volatility and market expectations:
  - Implied volatility in the range of 30 percent.
  - Market implied distributions were right-skewed (higher probabilities for further price increases).
  - Options activity indicated expectations of large upward jumps (price and volume of options at strikes in the range of US$75–US$85/bl).

### II. Empirical aspects of futures oil prices during 2002–06
- Time-series properties:
  - Augmented Dickey-Fuller test on level price (Null: St has a unit root):
    - Augmented Dickey-Fuller test statistic = -0.52; probability value = 0.88.
    - Test critical values: 1% level (-3.44); 5% level (-2.86); 10% level (-2.57).
  - Augmented Dickey-Fuller test on first differences (Null: ΔSt has a unit root):
    - Augmented Dickey-Fuller test statistic = -35.98; probability value = 0.00.
    - Test critical values: 1% level (-3.44); 5% level (-2.86); 10% level (-2.57).
  - Oil price modeled as AR2:
    - AR2 specification: S_t = 1.12 S_{t-1} - 0.31 S_{t-2} + 0.98? (as presented in source notation)
    - Coefficients shown in source: 12 31.32.30.98 0.930.070.09 (presented in table form)
    - R^2 = 0.99, DW = 2.05.
  - Stationarity: Levels possess a unit root; changes ΔS_t are stationary.

- Macroeconomic/market fundamentals during sample:
  - World crude oil supply: rigid at about 84 million barrels per day (mbd) for most of sample.
  - World economy growth: about 4–5 percent in real terms during 2002–2006.
  - Demand price elasticity range: -0.01 to -0.04 (reference to IMF WP/06/62).
  - Short-run income elasticity estimates: between 0.2 and 0.4.
  - Implication: price inelastic demand and rigid supply imply small excess demand requires large price changes to clear markets.

### Distributional properties of daily log-price returns
- Return definition: x_t = log S_t - log S_{t-1}.
- Empirical distribution features:
  - Mode around 1–2 percent.
  - Daily changes in the range of 5–7 percent were not uncommon.
  - Frequency of jumps exceeding ±3 percent estimated at 23 percent.
  - Descriptive statistics:
    - mean = 0.116
    - standard deviation = 2.29 (annualized to 36.3 percent)
    - skewness = -0.39
    - kurtosis = 4.79
    - Jarque-Bera normality statistic = 179.3; probability-value = 0.0
  - Interpretation:
    - Left-skewed distribution: downward jumps of smaller size more frequent than upward jumps of larger size, but positive mean implies larger upward jumps outweigh smaller downward ones.
    - Fat tails: large jumps occur more frequently than under normality.
  - Conclusion: presence of skewness and kurtosis suggests modeling via jump-diffusion or general Levy processes.

### B. Oil price time-varying volatility
- Two volatility measures analyzed:
  - Implied volatility from Brent futures options (August 2005–June 2006; source: Bloomberg).
    - Average implied volatility ~30 percent.
    - Volatility often surged to 34–35 percent.
  - GARCH(1,1) estimated volatility (January 2, 2002–July 7, 2006).
    - Return definition used: x_t = log S_t - log S_{t-1}.
    - GARCH mean volatility estimated at 43 percent.
    - Observed features: high volatility, volatility clustering, volatility rising during large price shocks and receding during price retreats.
- Market implications:
  - High volatility indicates large uncertainty and sensitivity to news and shocks.
  - Volatility clustering increases speculative demand for futures, which can further raise volatility.

### III. Crude oil price as a Merton jump-diffusion (J-D) process — model structure
- Continuous-time SDE for spot price S_t (J-D specification given in source):
  - dS_t = S_t [α dt + σ dB_t + (e^{J_t} - 1) dN_t]
  - Notation and assumptions:
    - α = instantaneous return; σ^2 = instantaneous variance.
    - B_t = standard Brownian motion with dB_t ~ (0, dt).
    - N_t = Poisson counter with intensity λ; arrivals independent of Brownian motion.
    - Jump size J_t iid ~ N(β, δ^2).
    - When a jump occurs, S_t = S_{t-} e^{J_t}; percentage change = e^{J_t} - 1.
- Log-price dynamics (using X_t = log S_t and Ito’s lemma):
  - dX_t = (α - 1/2 σ^2) dt + σ dB_t + ∑_{i=1}^{N_t} J_i
  - Drift parameter μ defined in source: μ = α - 1/2 σ^2.
  - Parameter vector: θ = (μ, σ^2, λ, β, δ^2).
- Discrete-time formulation (Δ = 1 day in presentation):
  - x_t = X_t - X_{t-1} = μ + diffusion component + sum of jumps over interval.
  - Density of x_t is a convolution of Gaussian diffusion and compound Poisson jump components:
    - Probability density (Δ = 1) expressed as an infinite sum over n = 0,1,2,... with Poisson weights and normal densities (as presented in source).
- Motivations for jumps:
  - Presence of skewness and kurtosis in returns.
  - Short-term implied volatility skews across strikes in option markets.
  - References: Merton (1976); Bakshi et al. (1997); Bates (1996).

### IV. Alternative estimation approaches (mentioned)
- Methods identified for estimating the J-D model:
  - Maximum Likelihood.
  - Method of Cumulants.
  - Method of Characteristic Function.
- (Details on estimation procedures and parameter estimates appear in subsequent sections of the source document.)

*Source: _wp06299 - 1. Time-Series Properties of the Oil Price (IMF working paper chapter text provided).*

### 1.  The maximum likelihood method: Let

### _wp06299 - 1.  The maximum likelihood method: Let

### Maximum likelihood (ML) method
- Log-likelihood for an observed sample of T log returns xt, t = 1,...,T is given (as in the source) by the displayed expression for Lx(·) involving μ, β, λ, σ, δ and double summations.
- Application issues:
  - Difficulties arise from identification of the jump parameter and instability of parameter estimates.
  - Computational tractability, un-boundedness over the parameter space, and instability of parameters have led researchers to alternative methods (method of moments, cumulants, CF-based methods).
- Historical applications:
  - Ball and Torous (1983) applied ML truncating the number of jumps at n = 15.
  - Ball and Torous (1985) and Jorion (1988) applied ML assuming a Bernoulli process for the jump component.
  - ML estimates achieve the lower bound for the Cramer-Rao efficiency criterion but practical problems motivate alternatives.

### Method of cumulants
- Characteristic function (CF) of Xt: φX(u) = E[exp(iuXt)] = ∫ exp(iuX) fX(X) dX.
- Cumulants κn are coefficients in the power series expansion of ln φX(u).
- CF for the jump-diffusion (J-D) process given in the source yields first four cumulants:
  - κ1 = μ + λβ
  - κ2 = σ2 + λδ2 + λβ2
  - κ3 = λβ3 + 3λβδ2
  - κ4 = λ(β4 + 6β2δ2 + 3δ4)
- Press (1967) imposed μ = 0 and derived closed-form relations for β̂, λ̂, δ̂, σ̂ in terms of sample cumulants; Press’ estimates were often wrong-signed and not plausible.
- Beckers (1981) set β = 0 and used sixth-order cumulants to derive estimators; improved over Press but anomalies remained.
- Ball and Torous (1983) used a Bernoulli jump process and Beckers’ restriction β = 0 to derive a system of cumulant equations and closed-form expressions for μ̂, λ̂, σ̂2, δ̂2 (explicit forms given in the source).

### Method of the characteristic function (CF) and empirical CF (ECF)
- CF and the corresponding probability density contain equivalent information; the CF is often available in closed form even when transition densities are not.
- Method of moments/GMM and ECF procedures:
  - Moments of any order n can be computed from derivatives of the CF: En[Xn] = i−n (d nφ(u)/du n)|u=0.
  - ECF: empirical φn(u) = (1/T) ∑_{j=1}^T exp(iu xj) = (1/T) ∑ (cos(uxj) + i sin(uxj)).
  - Estimator θ̂ = argminφn − φθ' W (φn − φθ), where W is positive semi-definite.
  - ECF is essentially equivalent to GMM; with sufficiently large sample and fine grid in u, ECF estimators can reach arbitrarily high asymptotic efficiency (Feuerverger 1990).
  - Over-identification (r > l) allows tests of over-identifying restrictions (Hansen (1982) type tests).

### Empirical estimation results (Brent futures daily sample)
- Three estimation approaches applied: ML (Bernoulli jump process), cumulants (various restrictions), ECF (unrestricted and with λ = 0.23).
- Common empirical findings across methods:
  - Dynamics influenced by both diffusion and jump components; jump component is dominant.
  - Jump component has much higher variance than the diffusion component, consistent with excess kurtosis.
  - Mean of jump size tends to be negative, consistent with negative skewness of returns.
  - Diffusion drift is high, consistent with observed upward trend in crude oil prices.
  - Jump dynamics: high intensity and large variance indicate frequent large discontinuities; markets often out-of-equilibrium.

- Selected parameter estimates (as reported in the source text):
  - Maximum Likelihood (Bernoulli jump process, unrestricted or with λ restricted in variants):
    - μ̂ = 0.23 (paragraph text)
    - 2σ̂ = 4.46
    - 2δ̂ = 4.47
    - λ̂ = 0.59
    - β̂ = −1.12
  - Method of cumulants (with λ = 0.23):
    - μ̂ = 0.32
    - 2σ̂ = 1.81
    - 2δ̂ = 2.78
    - β̂ = −0.85
  - Press (1967) restriction μ = 0 produced an implausible estimate:
    - 2δ̂ = −13.88
  - Beckers (1981) restriction β = 0:
    - μ̂ = 0.12
    - 2σ̂ = 3.34
    - 2δ̂ = 8.62
    - λ̂ = 0.22
  - ECF (unrestricted):
    - μ̂ = 0.57
    - 2σ̂ = 0.54
    - 2δ̂ = 0.96
    - λ̂ = 4.37
    - β̂ = −0.10
  - ECF with λ = 0.23:
    - μ̂ = 0.27
    - 2σ̂ = 3.34
    - 2δ̂ = 6.97
    - β̂ = −52

- Interpretation provided in the source:
  - Estimates concordant: oil price process dominated by jump process with large discontinuities at high intensity.
  - Negative mean of jump component interpreted as short-lived downward adjustments in demand after large upward jumps in price.
  - Large positive diffusion drift reflects a strong income effect that pushes oil prices upward, dominating the negative mean of jump component for daily data.
  - Economic explanation: World demand highly elastic to world income (which expanded 4–5 percent per year during the sample), highly inelastic to price; supply rigid — small excess demand leads to large price variation; income effect dominates price effect.

### Variance-Gamma (VG) Lévy process — motivation and definition
- Limitations of J-D model:
  - Does not capture time-varying (stochastic) volatility.
  - Models finite large jumps but cannot capture infinite small jumps analogous to diffusion.
- VG Lévy process proposed to capture stochastic volatility and small frequent jumps while retaining tractable CF.
- Definition (as in source):
  - VG process XVG t = α Gt + σ B_{Gt}, where {Gt, t ≥ 0} is a gamma process (mean rate 1/υ and variance rate 1/υ), B is Brownian motion, σ > 0, α ∈ R.
  - CF of VG: φVG t(u; σ, υ, α) = [1 − iαuυ + (1/2)σ2u2υ]^{-t/υ} (equivalently as displayed in the source).
- VG parameters and moments (under (σ, υ, α)):
  - Mean = α
  - Variance = σ2υ + α2
  - Skewness = (2αυ + 3α(σ2υ + α2)?) — exact expression provided in source (preserve source wording): skewness = (2 α υ + ... as in the source)
  - Kurtosis = expression in source: kurtosis = 4222 3(1   2()   )υυσσ υα−+− + (as displayed)
- Role of parameters:
  - α controls skewness (0α< implies negative skewness).
  - υ controls kurtosis and frequency of jumps (larger υ → more frequent jumps, fatter tails).

*Source: _wp06299 - 1.  The maximum likelihood method: Let (PDF chapter/section)*

### 2.  Estimation of the Variance-Gamma process

### 2.  Estimation of the Variance-Gamma process

### Model specification and characteristic function
- Price model: St = S0 exp(μt + Xt), where Xt is a Variance-Gamma (VG) process.
- Log return: Δxt = log(St) − log(St−1) = μ + Xt − Xt−1.
- Characteristic function (CF) of the log price return (notation preserved from source):  
  [][][]1 22 11 1 [exp()]exp(    ()exp()exp()exp()(1) 2 t EiuxEiuXiuEiuXiuiuu υ μ μ μαυσυ − =+==−+.
- For the VG model under the risk-neutral measure (asset price 0 exp[(rt + ω(α,σ,υ)) + Xt(σ,α,υ)]): Madan et al. (1998) CF for log(St):  
  22 0 1 (  )    exp[ln(    )    () ](1) 2 t t uSrtiuu υ φωαυσυ − =++−+.

### Parameter estimation (ECF results and interpretation)
- Estimation method: Empirical Characteristic Function (ECF) approach (see Table 3 in source).
- Textual summary of ECF findings:
  - Estimated parameters of the VG process were stable and statistically significant, corroborating the J-D process results.
  - Crude oil prices exhibited a high drift coefficient, high volatility, frequent and large jumps, and skewness.
  - The drift coefficient asserted a strong upward pulling force that kept reigniting oil prices.
  - Volatility parameter was high and significant, indicating high uncertainties about future price movements.
  - Parameter υ (tail fatness) was high and significant, implying frequent and large jumps.
  - Parameter α (skewness) was negative and significant, showing left-skewness in the VG distribution; interpretation: large positive jumps were typically followed by a cooling-off period with small/frequent negative jumps, while higher income demand elasticity (vs. price elasticity) meant faster world economic growth pushed demand upward.
- Explicit parameter values reported in paragraph text (preserve exact appearance):
  - ˆ μ = 0.09
  - ˆ σ = 1.66
  - ˆ υ = 0.80
  - ˆ α = −0.36  (textual line: "α, which controls for skewness, estimated at ˆ 0.36 α=−, was negative and significant")
- Table 3 parameter values (as presented in the table):
  - Drift μ: 0.09 (t = 3.52)
  - Drift α: −0.146 (t = −2.54)
  - Volatility σ: 1.66 (t = 80.4)
  - Variance of VG υ: 0.804 (t = 52.3)
  - Note: table footnote — "1/ Using ECF method."

### Option pricing using characteristic functions (methodology and formulas)
- Option pricing in asset price space: option value f(t,S t) = EQt[e−∫tT r(s) ds g(ST) | St = S] (source presents the convolution of discounted payoff with state price density).
- CF definition for log-price under risk-neutral measure: φTt(u) = ∫ p(ST | St) eiu log(ST) dST.
- Two Fourier-space pricing approaches:
  - Heston (1993): numerical inversion of the CF (singularity at u = 0 problematic).
  - Carr and Madan (1999): numerical inversion of the Fourier transform of the option value (damped option price approach).
- Carr-Madan transform notation (preserved):
  - Damped option price: cT(k) ≡ exp(a k) CT(k) for a > 0.
  - Fourier transform of damped option price: ψT(u) = ∫ eiu k cT(k) dk.
  - Relation between ψT(u) and φT(u) (as given):  
    22 ((1)) () (21) rT T T euai u aauia u φ ψ − −+ = +− + + .
- FFT inversion to recover option prices:
  - Inverse formula (preserved notation):  
    ( )() 1 ˆ 2 iuS fSte f u du π ∞ − −∞ = ∫.
  - Discrete N-point approximation with grid spacing Δ and quadrature weights wj as described; Simpson’s or trapezoidal rule can be used; grid considerations: Δ small, NΔ large enough to assume φ(u) ≈ 0 for u ≥ NΔ.

### Risk-neutral calibration (inverse problem) and density forecast for crude oil
- Inverse problem: back out risk-neutral VG parameters θ = (σ, α, υ) by minimizing quadratic pricing error over N observed option prices, subject to put-call parity constraint.
- Objective (preserved formula): θ̂ = arg min(θ) (1/N) Σj=1..N [CTK* (·,θ) − CTK(·)]^2 under put-call parity: SP + CTK* − CTK = Ke−rT.
- Regularized minimization with penalty parameter h > 0 (preserved formula):  
  θ̂ = arg min (1/N) Σj [CTK* − CTK]^2 + h [e−rT S0 + CTK − PTK − Ke−rT]^2.
- Data and calibration specifics (preserve exact figures and dates):
  - Data date: July 21, 2006.
  - Options: call and put futures options contracts maturing end-September 2006.
  - Risk-free interest rate used (three-month U.S. Treasury bill rate): 4.965.
  - Crude futures price: US$74.43/bl.
- Constrained minimization output (risk-neutral triplet reported):
  - ˆ σ = 1.72
  - ˆ υ = 1.12
  - ˆ α = 0.37
- Interpretation of implied risk-neutral distribution (July 21, 2006 for end-September 2006 maturity):
  - Market participants expected no short-term change in underlying fundamentals.
  - Expected high volatility: ˆ σ = 1.72.
  - Dominated by a jump process: ˆ υ = 1.12.
  - Right-skewed risk-neutral distribution (ˆ α = 0.37), indicating higher probability mass for prices rising above the futures price than falling below it.

### Main conclusions and implications
- Empirical period studied: January 2, 2002 – July 7, 2006 (conclusions drawn across this sample).
- Core empirical findings:
  - Dynamics dominated by frequent jumps causing markets to be often out-of-equilibrium.
  - Strong positive drift persisted, continually pushing prices upward despite retreats after major upward jumps.
  - High volatility made oil prices very sensitive to small shocks and news.
  - Both J-D and VG specifications produced consistent findings aligned with oil-market fundamentals and world economic conditions.
  - Faster world economic growth and expansionary monetary policies expanded crude demand; given price inelasticities of demand and supply, small excess demand/supply led to large price adjustments — explaining high jump intensity and upward pressure.
- Policy and practical relevance:
  - Results are relevant for forecasting, risk management, derivatives pricing, and gauging market sentiment.
  - Findings help establish the stochastic-process nature of oil prices and link process parameters to underlying fundamentals — implying parameter changes require fundamental shifts.
  - Relevance for the Fund in monitoring oil markets and seeking policies to stabilize these markets.

### Annex summary: Method of cumulants (key relations preserved)
- Moment generating function M(u) = E[e uX ] = ∫ fX(x) e u x dx.
- Cumulant-generating function: log M(u) = Σn=0..∞ κn u n / n!.
- Relation between raw moments mj = E[X^j] and cumulants κn via recursion (preserved form):  
  1 1 1 1 n nn n n j j n mm j κ κ − − − = − ⎛⎞ =− ⎜⎟ − ⎝⎠ Σ.
- Selected relations preserved:
  - m1 = κ1
  - m2 = κ2 + κ1^2
  - Central moment relations: first central moment 0, second central moment = κ2, third central moment = κ3, fourth central moment = κ4 + 3 κ2^2.
- Kurtosis measure preserved: kurtX = E[X^4] − 3 (E[X^2])^2 (source formula expressed as () () () () 2 42 3kurt   XE   XE   X=−).

*Source: _wp06299 - 2.  Estimation of the Variance-Gamma process (IMF working paper PDF content provided in prompt).*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06299.pdf_