state variables. Spot prices are an example of a state variable, as they are not
directly observable. The spot prices are generated by a discrete time version of
the stochastic process used, called transition equation. The Kalman Filter is
then applied recursively to calculate the optimal estimator for the state vector
at time t using the information available at t. We shall treat this in detail in
section 2.1.
As mentioned by Schwartz (1997), the main difficulty in the empirical im-
plementation of commodity price models is that frequently the state variables,
such as the spot price for example, are not directly observable. The finan-
cial literature traditionally estimates the parameters of a model with hidden
factors by rewriting it in state space form and using the Kalman Filter.
Barlow, Gusev and Lai (2004) provide a good explanation of the Kalman’s
Filter application, and use it to calibrate three different models for electricity
prices. Schwartz (1997) also exposes the construction of the Kalman Filter in
a simple, but complete way.
In this work, we propose a methodology to recover spot prices from futures
without the use of Kalman Filter. We use maximum likelihood and nonlinear
least squares in a two-factor affine model based on Schwartz and Smith (2000).
The model we use describes commodity spot prices as a sum of short term
deviations over a long term pattern. When performing a Principal Compo-
nent Analysis in the data set of natural gas prices, for instance, we find that
the behavior of the two most important eigenvectors are very much like the
simulated paths for the short term deviations and long term dynamics con-
sidered by Schwartz and Smith (2000). It is important to note also that our
calibration method can be applied to other energy commodities other than
natural gas.
Our estimation procedure resembles to some extent the work done in Hik-
spoors and Jaimungal (2007). We use a nonlinear least-squares optimization
(NLLS), as in Hikspoors and Jaimungal (2007), to recover the hidden short
and long term factors from the futures data. However, we used a maximum
likelihood estimation in order to recover the model’s parameters, instead of
the further NLLS minimization used in Hikspoors and Jaimungal (2007).
The main advantage of our calibration procedure is its robustness. As we
show in Chapter 4, the model converges quickly for different initial guesses and
even with large perturbations. Besides that, our algorithm is computationally
cheaper than the Kalman Filter. It does not need to impose restrictions in
order to decrease the number of estimated parameters and it is not dependent
of a wise initial-guess selection in order to obtain convergence.
This work shall be divided as follows: Chapter 2 explores the literature
concerning commodity modeling and estimation via Kalman Filter. We detail
3