Page: 1 Carbon Phase II Density Applications using General Scientific Stochastic Volatility Models by Professor Per B Solibakke a Nord Pool Application: Carbon Front December Future Contracts a) Department of Economics, Molde University College Workshop Molde, Session Monday 21/May/2012: 14:00am – 15:30am
58
Embed
Page: 1 Carbon Phase II Density Applications using General Scientific Stochastic Volatility Models by Professor Per B Solibakke a Nord Pool Application:
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page: 1
Carbon Phase II Density Applications
using General Scientific Stochastic Volatility Models
by
Professor Per B Solibakkea
Nord Pool Application: Carbon Front December Future Contracts
a) Department of Economics, Molde University College
1. The Front December Future Contracts NASDAQ OMX: phase II 2008-2012
No EUAs the theoretical spot-forward relationship does not exist
Price dynamics are depending on total emissions (extreme-dynamics)
EUA options have carbon December futures as underlying instrument
Page: 2
2. The dynamics of the forward rates are directly specified
The HJM-approach adopted to modelling forward- and futures prices in commodity markets
Alternatively, we model only those contracts that are traded, resembling swap and LIBOR models in the interest rate market ( also known as market models). Construct the dynamics of traded contracts matching the observed volatility term structure
The EUA options market on carbon contract are rather thin (daily prices are however reported), we estimate the model on the future prices themselves. Black-76 / MCMC simulations
3. Stochastic Model Specification: Estimation, Assessment and Inference
4. Forecasting unconditional Futures Moments, and Risk Management and Asset Allocation measures
5. Forecasting conditional Futures Moments
i. One-step-ahead Conditional Mean (expectations) and Standard Deviation
ii. Conditional one-step-ahead Risk Management and Asset Allocation measures
iii. Volatility/Particle filtering for Option pricing
iv. Multi-step-ahead Mean and Volatility Dynamics
v. Mean Reversion and Volatility Persistence measures
6. The EMH case of CARBON Futures/Options for Commodity Markets
Page: 3
Page: 4
The Carbon NASDAQ OMX commodity market
NASDAQ OMX commodities provide market access to one of Europe’s leading carbon markets.
350 market members from 18 countries covering a wide range of energy producers, consumers and financial institutions.
Members can trade cash-settlement derivatives contracts in the Nordic, German, Dutch and UK power markets with futures, forward, option and CfD contracts up to six years’ duration including contracts for days, weeks, months, quarters and years.
The reference price for the power derivatives is the underlying day-ahead price as published by Nord Pool spot (Nordics), the EEX (Germany), APX ENDEX (the Netherlands), and N2EX (UK).
1. Projection: The Score generator (A Statistical Model) establish moments: the Mean (AR-model) the Latent Volatility ((G)ARCH-model) Hermite Polynomials for non-normal distribution features
2. Estimation: The Scientific Model – A Stochastic Volatility Model
where z1t , z2t and (z3t ) are iid Gaussian random variables. The parameter vector is:
Page: 5
The General Scientific Model methodology (GSM):
2
0 1 1 0 1, 2, 1
1, 0 1 1, 1 0 2
2, 0 1 2, 1 0 3
1 1
22 1 1 1 1 2
22 2 2
3 2 2 1 3 2 1 1 2 3 2 1 1 3
exp( )
1
/ 1 1 / 1
t t t t t
t t t
t t t
t t
t t t
t t t t
y a a y a u
b b b u
c c c u
u z
u s r z r z
u s r z r r r r z r r r r r z
0 1 2 0 1 0 1 1 2 1 2 3, , , , , , , , , , ,a a a b b c c s s r r r
1 10 10 12 2 13 3 1
2 22 2 2
3 33 3 3
exp( )t t t t
t t t
t t t
dU dt U U dW
dU U dt dW
dU U dt dW
SDE:
A vector SDE with two stochastic volatility factors.
Andersen and Lund (1997): Short rate volatilitySolibakke, P.B (2001): SV model for Thinly Traded Equity MarketsChernov and Ghysel (2002): Option pricing under Stochastic VolatilityDai & Singleton (2000) and Ahn et al. (2002): Affine and quadratic term structure modelsAndersen et al. (2002): SV jump diffusions for equity returnsBansal and Zhou (2002): Term structure models with regime-shiftsGallant & Tauchen (2010): Simulated Score Methods and Indirect Inference for Continuous-time Models
3. Re-projection and Post-estimation analysis:
MCMC simulation for Risk Management and Asset allocation Conditional one-step-ahead mean and volatility densities. Forecasting volatility conditional on the past observed data; and/or extracting volatility given the full data series (particle filtering/option pricing) The conditional volatility function, multi-step-ahead mean and volatility and mean/volatility persistence. Other extensions for specific applications.
Early references are: Kim et al. (1998), Jones (2001), Eraker (2001), Elerian et al. (2001), Roberts & Stamer (2001) and Durham (2003).
A successful approach for diffusion estimation was developed via a novel extension to the Simulated Method of Moments of Duffie & Singleton (1993). Gouriéroux et al. (1993) and Gallant & Tauchen (1996) propose to fit the moments of a discrete-time auxiliary model via simulations from the underlying continuous-time model of interest EMM/GSM
First, use an auxiliary (statistical) model with a tractable likelihood function and generous parameterization to ensure a good fit to all significant features of the time series.
Second, a very long sample is simulated from the continuous time model.
The underlying parameters are varied in order to produce the best possible fit to the quasi-score moment functions evaluated on the simulated data. Under appropriate regularity, the method provides asymptotically efficient inference for the continuous time parameter vector.
Simulated Score Methods and Indirect Inference for Continuous-time Models (some details):
The idea (Bansal et al., 1993, 1995 and Gallant & Lang, 1997; Gallant & Tauchen, 1997):
Use the expectation with respect to the structural model of the score function of an auxiliary model as the vector of moment conditions for GMM estimation.
The score function is the derivative of the logarithm of the density of the auxiliary model with respect to the parameters of the auxiliary model.
The moment conditions which are obtained by taking the expectations of the score depends directly upon the parameters of the auxiliary model and indirectly upon the parameters of the structural model through the dependence of expectation operator on the parameters of the structural model.
Replacing the parameters from the auxiliary model with their quasi-maximum likelihood estimates, leaves a random vector of moment conditions that depends only on the parameters of the structural model.
Page: 9
Simulated Score Methods and Indirect Inference for Continuous-time Models (some details): Estimation
Simulated Score Estimation:
Suppose that: is a reduced form model for observed time series, where xt-1 is the state vector of the observable process at time t-1 and yt is the observable process. Fitted by maximum likelihood we get an estimate of the average of the score of the data satisfies:
That is, the first-order condition of the optimization problem.
1( | , )t tf y x
n 1 1
,n
t t ty x
1
1
1/ / log | , 0
n
t t n
t
n f y x
Having a structural model (i.e. SV) we wish to estimate, we express the structural model as the transition density , where r is the parameter vector. It can be relatively easy to simulate the structural model and is the basic setup of simulated method of moments (Duffie and Singleton, 1993; Ingram and Lee, 1991).
The scientific model is now viewed as a sharp prior thatrestricts the -parameters to lie on the manifold
Main question: How do the results change as this prior is relaxed?That is: How does the marginal posterior distribution of a parameter or
functional of the statistical model change?
Distance from the manifold: where Aj is the scaling matrices.
Hence, we propose that the scientific model be assessed by plotting suitablemeasures of the location and scale of the posterior distribution of against , or, better sequential density plots.
.M H
| ,p y x
Page: 10
1,...
, minM j j jj G
d A
For a well fitting scientific model:The location measure should not move by a scientifically meaningful amount as k increases. The result indicates that the model fits and that the scale measure increases, indicating that the scientific model has empirical content.
Simulated Score Methods and Indirect Inference for Continuous-time Models (some details): Assessment
Page: 11
Simulated Score Methods and Indirect Inference for Continuous-time Models (some details): Re-projection / Post-Estimation Analysis
Elicit the dynamics of the implied conditional density for observables:
0 1 0 1ˆˆ | ,..., | ,..., ,L L np y y y p y y y
The unconditional expectations can be generated by a simulation:
0ˆ 0 0
ˆ... ,..., ,..., , ...Ln
L L n y yE g g y y p y y d d
ˆN
t t Ly
Now define: ˆ 0 1
arg maxˆ log | ,..., ,K n
K K LE f y y y
where is the score (SNP) density. 0 1| ,..., ,K Lf y y y
Let . Theorem 1 of Gallant and Long
(1997) states:
0 1 0 1ˆ ˆ| ,..., | ,..., ,K L K L Kf y y y f y y y
0 1 0 1ˆ ˆlim | ,..., | ,...,K L L
Kf y y y p y y y
We study the dynamics of by using as an approximation. p̂ ˆKf
Simulated Score Methods and Indirect Inference for Continuous-time Models (some details): Re-projection / Post-Estimation Analysis
Filtered volatility involves estimating an unobserved state variable conditional upon all past and present values. Denote v (unobserved) and y (contemporaneous and lagged observed variables) thus modified by v* and y*, respectively. Hence, simulate from the optimal structural
model and re-project to get
The price of an option (re-projected volatility) can now be calculated as:
Higher Understanding of the Carbon Futures Commodity Markets the Mean equations (drift, serial correlation, mean reversion) the Volatility equations (constant, persistence, asymmetry, multiple factors)
Models derived from scientific considerations and theory is always preferable Fundamentals of Stochastic Volatility Models Likelihood is not observable due to latent variables (volatility) The model is continuous but observed discretely (closing prices)
Bayesian Estimation Approach is credible Accepts prior information No growth conditions on model output or data Estimates of parameter uncertainty (distributions) is credible
Value-at-Risk / Expected Shortfall for Risk Management Stochastic Volatility models are well suited for simulation Using Simulation and Extreme Value Theory for VaR-/CVaR-Densities
Simulations and Greek Letters Calculations for Asset Allocation Direct path wise hedge parameter estimates MCMC superior to finite difference, which is biased and time-consuming
Re-projection for Simulations and Forecasting (conditional moments) Conditional Mean and Volatility forecasting Volatility Filtering and Pricing of Commodity Options Consequences of Shocks, Multiple-step ahead Dynamics and Persistence.
The Case against the Efficiency of Future Markets (EMH) Serial correlation in Mean and Volatility Price-Trend-Forecasting models and Risk premiums Predictability and Efficient use of Available Information
SV models has a simple structure and explain the major stylized facts. Moreover, market volatilities change so frequent that it is appropriate to model the volatility process by a random variable.
Note, that all model estimates are imperfect and we therefore has to interpret volatility as a latent variable (not traded) that can be modelled and predicted through its direct influence on the magnitude of returns.
Mainly three motivational factors:
1. Unpredictable event on day t; proportional to the number of events per day. (Taylor, 86)
2. Time deformation, trading clock runs at a different rate on different days; the clock often represented by transaction/trading volume (Clark, 73).
3. Approximation to diffusion process for a continuous time volatility variable; (Hull & White (1987)
4. A model of futures markets directly, without considering spot prices, usingHJM-type models. A general summary of the modelling approaches for forward curves can be found in Eydeland and Wolyniec (2003).
Matching the volatility term structure.
5. In order to obtain an option pricing formula the futures are modelled directly. Mean and volatility functions deriving prices of futures as portfolios.
Such models can price standardized options in the market. Moreover, the models can provide consistent prices for non-standard options.
6. Enhance market risk management, improve dynamic asset/portfolio pricing, improve market insights and credibility, making a variety of market
forecasts available, and improve scientific model building for commodity markets.
Carbon Density Applications for Front December Futures Contracts
Back to Overview
Scientific Models: Stochastic Volatility Model /Parameters (q)
Bayesian Estimation Results1. Several serial Bayesian runs establish the mode
We tune the scientific model until the posterior quits climbing and it looks like the mode has been reached: Then:
2. A final parallel run with 32 (8 cores *4 CPUs) CPUs and 240.000 MCMC simulations (Linux/Ubuntu 12.04 LTS & OPEN_MPI (Indiana University) parallell computing)
Page: 22
Carbon Front December General Scientific Model. Statistical Model SNP-11116000 - fit modelParameter values Scientific Model. Standard Parameters Non-linear-GARCH. Standard
Scientific Model: Model Assessment – the model concert test
Carbon front December k = 1, 10, 20 and 100 densities – reported.
Carbon Density Applications for Front December Futures Contracts
Page: 25
Scientific Model: The Stochastic Volatility Model: log-sci-mod-posterior
Log sci-mod-posterior (every 25th observation reported): Optimum is along this path!
c2(4)
-0.94841{0.81373}
Optimum:
Carbon Density Applications for Front December Futures Contracts
Page: 26
Scientific Model: Carbon q-paths and densities; 240.000 simulations
Carbon Density Applications for Front December Futures Contracts
Page: 27
Scientific Model: Stochastic Volatility
The chains look good. Rejection rates are:
The MCMC chain has found its mode.
A well fitted scientific SV model:The location measure is not moved by a scientifically meaningful amount as k increases. The result indicates that the model fits and that the scale measure increases, indicating that the scientific model has empirical content.
Reported Proportion Number of Proportion%-rejected Moved Rejects Accepted
The most predominant application is Option Pricing.
Step 1: A long simulation(i.e. 250 k): where v* is the unobserved volatility and y* the observed returns (incl. lags).
Step 2: Make a new projection to get the BIC optimal fK density with lags set generously long and non-linearity is available.
The result is the unconditional mean from the raw the simulations and the conditional volatility is the conditional mean from fK : The conditional volatility mean can be used to obtain an estimate of:
for the purpose of pricing an option (particle filtering).
Carbon Density Applications for Front December Futures Contracts
Risk Premium (non-diversifiable risk) calculations are based on information at t-1.
For contract i we obtain the risk premium Ri at t-1 as:
The risk premium is interpreted as non-diversifiable risk in the commodity market and is added as a constant to the Re-projected and the Black-76 average volatility for market comparisons.
Note that the risk premium does not imply arbitrage opportunities within a market if risk is treated consistently. However, the risk premium may induce arbitrage opportunities between carbon markets (i.e. NASDAQ OMX and ECX (theice.com)).
Carbon Density Applications for Front December Futures Contracts
Market prices are raw data at close for a specific date.
Re-projected Volatility and option prices:
(or any other more complex function)
N >= 250.000 conditional moment observations from re-projection step (see page 43 in this presentation) . The density is adjusted for the risk premium for contract i.
The Black´76 prices are calculated from an average of the re-projected volatility adjusted for the risk premium at t-1.
Carbon Density Applications for Front December Futures Contracts
Finally, we calculate the mean relative error and mean absolute error (calls):
where n is the number of strike contracts at time t, CMi is the model call
price and Ci is the observed market price (for put options P replaces C).
The MRE statistic measure the average relative biases of the model option prices, while the MARE statistic measures the dispersion of the relative biases of the model prices. The difference between MRE and MARE suggests the direction of the bias of the model prices, namely when MRE and MARE are of the same absolute values, it suggests that the model systematically misprices the options to the same direction as the sign of MRE, while when MARE is much larger than MRE in absolute magnitude, it suggests that the model is inaccurate in pricing options but the mispricing is less systematic.
Carbon Density Applications for Front December Futures Contracts
Scientific Model: The Stochastic Volatility Model.
Page: 50
The NASDAQ OMX market: The Re-projected Volatility Model Option Prices
MRE/MARE relative pricing errors (no. of contracts Sep-11 to Dec-12).
Moneyness Volatility Days to Maturityx=ln (S/K) Models < 30 dager 30 - 180 days 180 - 360 days > 360 days Overall
MRE MARE MRE MARE MRE MARE MRE MARE MRE MAREReprojected -0.2631 0.2631 -0.2020 0.2020 -0.1417 0.1417 -0.2023 0.2023
Option Prices are strongly influenced by risk premiums. The NASDAQ OMX market and the re-projected model show systematic option (mis-)pricing.
Page: 57
Carbon Density Applications for Front December Futures Contracts
Main Findings for CARBON Applications and Front December contracts:
Stochastic Volatility models give a deeper insight of price processes andthe stochastic flow of information interpretation
The Stochastic Volatility model and the statistical model seem to work well in concert (indirect estimation)
The MC chains look good and rejection is acceptable giving a reliable and viable stochastic volatility model
The SV-model results induce serial correlation in mean and volatility, persistence and negative asymmetry. One volatility factor is slowly moving while the second is quite choppy.
Risk management procedures are available from Stochastic Volatility models and Extreme Value Theory (VaR/CVaR and Greek letters)
Conditional moments, particle filtering and volatility variance functionsinterpret asymmetry, pricing options and evaluates shocks.
Option Prices can be generated from re-projected volatility for any maturity/ complexity. Non-diversifiable risk implies volatility risk premiums.
Imperfect tracking (incomplete markets) suggest that simulation is a well-suited methodology for derivative pricing.