Energy Pricing

Mean Reverting
Processes – Energy Price Processes
Used For Derivatives Pricing & Risk Management
This is the second article in a 3-part series exploring the main stochastic price processes
used to model energy spot and forward prices for derivatives valuation and risk manage-
ment. The first article in this series focused on the “random walk” assumption charac-
terised by the most popular of price processes, Geometric Brownian motion, and the Black-
Scholes option pricing model that is based on it. In this second article we shift our atten-
tion to the Mean Reverting Process, which incorporates the tendency of energy prices to
gravitate towards a "normal" equilibrium price level that is usually governed by the cost of
production and level of demand. By CARLOS BLANCO & DAVID SORONOW.

WHY DO we need to incorporate mean reversion when a random fashion. However, the size of his stumble is bound-
modelling energy prices? Suppose we observe that electric- ed by the length of the leash and the direction of his stride
ity prices jump from $30/MWh to $150/MWh due to an unex- tends towards Rusty’s position. When the drunk staggers
pected event (e.g. plant outages, transmission constraints, away from Rusty, he will eventually be pulled back, and in
heat wave, etc.). Most market practitioners would agree that due course will follow the path leading to his house.
it is highly probable that prices will eventually return to their
average level once the cause of the jump goes away. For sim- Random Walk With Mean Reversion: Mathematically
ilar reasons if the price of a barrel of WTI falls to US$7 due Mathematically, we capture the phenomena of mean rever-
to overproduction we would expect the price to eventually sion with a modification to the random walk assumption.

(S* —St) εt
rise as producers decrease supply. These expectations are
intuitive in nature and are supported by our observations of St+1 —St = α + σ
energy spot price behaviour.
These two simple examples illustrate the limitations of Expected change in Mean Reversion Random
Geometric Brownian motion (GBM) when applied to energy price at t+1 and t Component Component
prices. In the electricity example above GBM would accept where:
the US$150/MWh price as a normal event and would pro-
S* is the mean reversion level or long run equilibrium price
ceed randomly from there (via a continuous diffusion
process) with no consideration of prior price levels (no mem-
St is the spot price

ory), and no greater probability of returning to the average α is the mean reversion rate
price level. This result is clearly at odds with market reality, σ is the volatility
and provides the impetus for the modelling of more complex
mean reverting price processes.
ε is the random shock to price from t to t+1

Notice from this equation that the mean reversion compo-

A Random Walk With Mean Reversion nent or ‘drift’ term is governed by the distance between the
Strictly speaking, the ‘random walk’ used to model prices current price and the mean reversion level as well as by the
under GBM is based on the assumption that price changes mean reversion rate. If the spot price is below the mean
are independent of one another. In other words, the his- reversion level, the mean reversion component will be posi-
torical path the price followed to achieve its current price tive, resulting in an upward influence on the spot price.
is irrelevant for predicting the future price path (prices fol- Alternatively, if the spot price is above the mean reversion
low a Markov process). Mean reversion can be thought of level, the mean reversion component will be negative, thus
as a modification of the random walk, where price changes exerting a downward influence on the spot price. Over time,
are not completely independent of one another but rather this results in a price path that drifts towards the mean rever-
are related. sion level, at a speed determined by the mean reversion rate.
An analogy is often made that the random walk process is The following graphs illustrate several simulated price
like a stumbling drunk after leaving the bar; the direction paths assuming GBM vs. GBM with mean reversion. Notice
and size of his stumble are unknown and do not depend on that the GBM process produces prices that can reach unre-
his previous stride. Mean reversion can also be explained alistic levels for long periods of time. Unlike pure GBM, the
with a similar comparison. Now imagine that the drunk is mean reverting GBM price process ensures prices gravitate
being guided home by his German Shepherd dog, “Rusty”. over time toward the mean reversion price levels. Later in
We want to determine the evolution of the distance between this article we will explore the implications of this property
the drunk and the dog. The drunk will still stumble along in on derivative pricing and risk management.

June 2001 * Commodities Now 68

 Energy Pricing

Gemeotric Brownian Motion: Gemeotric Brownian Motion With Mean Reversion:

Sample Price Paths Sample Price Paths
4000 200
180
3500
160
3000
140
2500 120

Price
Price

2000 100
80
1500
60
1000
40
500 20
0 0

May 2001

Jun 2001

Sep 2001

Dec 2001
Jul 2001

Aug 2001

Oct 2001

Nov 2001

Jan 2002
May 2001

Jun 2001

Sep 2001

Dec 2001
Jul 2001

Aug 2001

Oct 2001

Nov 2001

Jan 2002
Terminal price Terminal price
distribution distribution

Source: @Energy. Financial Engineering Associates

FEA’s @Energy’s Mean Reverting Model (Extracted from @Energy User’s Guide)
Model Parameters: In FEA’s @Energy, the following equation governs the evolution of St in the mean reverting model
The Mean Reversion Level & Mean Reversion Rate without jumps:

( F (t 0,t ) )] dt + σt dwt
One of the great advantages of GBM models is the ease of S
estimating input parameters. The most difficult parameter d log St = a θt – log [ (1)
to estimate is volatility: the expected future variability of
price over time. This is typically estimated using either his- where St is the ‘spot’ price of the commodity at time t, wt is a standard Brownian motion, a >0
torical volatility, or implied volatilities from current option is a constant mean reversion rate (speed), σ t is a (possibly) time-dependent local volatility,
quotes. In contrast, mean reversion models require the esti- F (0,t) is the forward price at time zero for delivery of the commodity at time t , and log denotes
mation of additional unknown parameters such as the mean the natural log.
reversion level (the long-run equilibrium price), and the
mean reversion rate (the speed at which prices revert). The time-dependent function θt is internally determined by requiring that the expected value of
Depending on the exact formulation of the Mean Reversion S at time t (determined at time zero), match the input forward price F (0,t). This guarantees
model, these parameters can often be extracted from current that the process used is consistent with market prices of forward and futures contracts, and is
forward/futures prices and historical spot price series. A required for arbitrage-free pricing of the derivatives valued by @ENERGY.
common approach is to assume the current forward/futures
prices represent time dependent mean reversion price levels. Equation (1) provides a way to compute the evolution of the full forward price curve by using infor-
The premise here is that forward/futures prices are the mar- mation about:
ket’s "best guess" (unbiased estimate) of future spot prices. • The initial forward price curve F (0,t) for t > 0.
This approach has the added advantage of incorporating sea- • A mean reversion rate.
sonality in the price process since the term structure of for- • An estimate of the volatility of S as a function of time (technically, an instantaneous forward
ward prices usually has seasonal characteristics embedded volatility curve).
within it. The mean reversion rate is always greater than, or equal to, zero and higher numbers correspond to
There are several techniques for estimating the mean rever- faster mean reversion. The mean reversion parameter, a, represents the annualised rate at which the
sion rate. Linear regression can be used to relate historical underlying short-term price returns to its expected value. Hence the inverse of a gives the actual time
price changes to historical prices. More recently, complex cali- scale over which mean reversion occurs. For example, a mean-reversion rate of 2 corresponds to a com-
bration techniques that fit model parameters to historical spot modity whose price reverts to its expected value over the course of six months (i.e., 1 / 2 years).
price data have been devel-
oped. Although it requires extra
FEA – Random Price Simulator–Mean Reverting Lognormal Process
effort to determine these Market Inputs Value
25
parameters, the mean reversion Forward Price 20
24
Volatility 44%
model provides a much more Spot Price 23
22
accurate depiction of market Interpolation 1
Futures Contract 21
Price

reality, especially for options 20

pricing and risk management Simulation Inputs Value 19
Simulation sate date 01/01/02
18
purposes. Simulation end date 01/01/10
17
The speed of mean reversion Sampling Frequency m
16
Seed*
depends on several factors: the Simul* 10000 15
Sep 2004

Sep 2008

Jan 2010
Jan 2002
May 2002
Sep 2002
Jan 2003

Jan 2004
May 2003
Sep 2003

May 2004

Jan 2005
May 2005
Sep 2005
Jan 2006
May 2006
Sep 2006
Jan 2007
May 2007
Sep 2007
Jan 2008
May 2008

Jan 2009
May 2009
Sep 2009

commodity being analysed, the Min-Max*

delivery provisions associated Model Inputs Value
with the commodity, and so on. Model
Mean-Revision rate 24
In electricity markets, it is com- Cap on Price
mon to observe sudden price Floor on Price
Yearly jump frequency 2
spikes with very fast mean Expected jump size 0.6
reversion to the previous price St. Dev. of jump 0

levels before the price jump. In Other Inputs

15.50

15.93

16.48

16.98

17.51

18.03

18.54

19.06

19.57

20.09

20.60

21.12

21.63

22.14

22.66

23.18

23.69

24.21

24.73

25.20

25.81

Natural Gas markets, the mean Value Date* 01/01/00

Interest rate 0.00%
reversion rate is considerably

69 Commodities Now * June 2001

 Energy Pricing

slower, but the volatilities for longer-dated contracts are usu- ing to purchase cogeneration plants found themselves out-
ally lower than the volatilities for the shorter-dated ones. In bid by competitors. In such cases, these competitors felt
oil markets, the mean reversion rate is thought to be longer confident in bidding higher because their grasp of the deriv-
term, and it can take months, or even years, for prices to atives viewpoint gave them a greater understanding of the
revert to their mean. value of the plant. The losing utilities had instead relied on
traditional NPV analysis.
Implications For Derivatives Valuation & Risk Management: For asset valuation and risk analysis, we can follow a two
Option Pricing step process. Firstly we decompose the Fixed Asset as a set
The mean reversion model has significant implications for of exposures to market factors. Our profit/loss each day is
option pricing, especially for American exercise options and going to be a function of the spot prices recorded that day,
certain exotic instruments such as barrier options and swing the power generated, load, etc. In the second step we simu-
options. late the different variables that determine our hypothetical
It is common practice for traders to estimate volatility by profit/loss for that day.
calculating the Black-Scholes model implied volatility from a For example, suppose that a generation facility produces
set of quoted European exercise option prices. These volatil- electricity by burning natural gas. From a derivatives per-
ities are then used to price less liquid exotic options such as spective, the operator has the option to transform gas into
barriers, spread options, and swing options using GBM electricity. In derivatives vernacular, the generation plant is
based option pricing models. However, many practitioners nothing more than a basket or “strip” of spread options,
may be unaware that this approach can produce erroneous since such choices can be repeatedly made during the facil-
results, especially in high price volatility environments such ity’s useful life. From a risk management point of view, the
as is the case in most power markets. operator has exposures to power and natural gas prices. He
Consider a natural gas call option that ‘knocks-out’ and will be short natural gas (because he needs natural gas to
becomes worthless if the price of natural gas rises above fire his generators), and he will be long power (because he
US$12/mmBtu at anytime during the life of the option. Using has the ability to produce electricity). But his exposure to
a GBM price process to model natural gas prices produces electricity, as shown by the power and natural gas option
price paths that result in a much higher probability of reach- deltas, wouldn’t necessarily be 100% of the output of the
ing the barrier level during the option life than a mean revert- plant since he can decide not to run the plant.
ing price process does. Option pricing models that use a In order to value this “derivative” it is essential that our
mean reverting price process ensure prices gravitate towards possible price scenarios are plausible. Unless we use the
the mean reversion level, thereby assigning less probability right stochastic process, our simulated profits could be
of touching the barrier during the life of the option. totally unrealistic. If we used a “random walk” process to
The graphs below illustrate simulated price paths and simulate electricity prices for the next 10 years, we would
resulting histograms for natural gas using GBM vs. GBM find that the simulated prices after the first few months
with mean reversion. The higher prices produced by the would not accurately represent their expected behaviour.
pure GBM method can be clearly seen. In this example, both Another important aspect is the “greek letters” or sensi-
price processes produce the same result for a European tivities of the derivative to the various components deter-
exercise option, but drastically different option prices for a mining its value. If we wrongly assume that prices follow a
barrier option. non-mean reverting process, we would find that the greek
measures become meaningless in a hedging context, espe-
Mean Reversion & Real Options Analysis cially for assets with very high volatilities. For example, if we
Another example of the dangers of not taking account of use GBM to value European options with very high volatili-
mean reversion comes from asset valuation. Many compa- ties, the delta converges to 1 for call options and –1 for put
nies are starting to treat energy related fixed assets as options. This is also the case for the deltas in Spread
derivative instruments using an approach known as “real Options where one of the assets has very high volatility. The
options analysis”. The downside of ignoring the derivatives deltas of the underlying assets (e.g. power and gas) con-
pricing approach has been demonstrated dramatically in US verge to 1 and –1 respectively, due to the erroneous choice of
power markets. Many utility companies who were attempt- the right price process to determine those price sensitivities.

Gemeotric Brownian Motion: Gemeotric Brownian Motion With Mean Reversion:

Sample Price Paths Sample Price Paths
20
10
18
9
16
8
14
7
12
6
Price

Price

10
5
8
4
6
3
4
2
2
1
0
0
May 2001

Jun 2001

Sep 2001

Dec 2001
Jul 2001

Aug 2001

Oct 2001

Nov 2001

Jan 2002

May 2001

Jun 2001

Sep 2001

Dec 2001
Jul 2001

Aug 2001

Oct 2001

Nov 2001

Jan 2002

Terminal price Terminal price

distribution distribution
Source: @Energy. Financial Engineering Associates

June 2001 * Commodities Now 70

If we use a mean reverting process, we would obtain more
accurate “deltas” or price sensitivities that would produce a
more effective hedge strategy.

Value at Risk & The “Square Root Of Time Rule”

Our final example of the problems with GBM concerns the
square root of time rule commonly used in Risk
Management to convert certain Value at Risk measures
between different holding periods. If prices follow a mean
reverting process, the “square root of time rule” would give
us very large volatility estimates. Note that in a mean revert-
ing process, the volatility has a limited effect on prices, and
due to the fact that after a shock, prices revert to their long
run levels, their long run variability does not grow propor-
tionately with time.

How To calculate Mean Reversion Rates, Mean

Reversion Levels & Volatilities In Excel
This practical example illustrates the problem with the
square root of time rule.
1. Choose a particular price series. Prices could be hourly,
daily, weekly, etc. Column A (rows 2–11) contains historical
prices for asset x over a ten-day period.
2. Calculate the standard deviation assuming that returns
are independent. Column B (rows 3–11) shows the daily
returns, denoted u. Use Excel’s built-in STDEV (sample stan-
dard deviation) function to calculate the daily volatility (cell
B13). Using the square-root-of-time rule, we annualise to
obtain s* (cell B15) assuming that time is measured in trad-
ing days and there are 250 trading days per year. In our
FEA
example, the annualised volatility would be equal to 265%,
(16.7% x √250 ).
3. Calculate the absolute price changes. Column D (rows
HPFC
3–11) shows the daily changes.
4. We can estimate the Mean Reversion rate in a relatively
simple and robust manner by regressing absolute price
changes (Column D) on the previous price levels (Column E).
5. Use the Excel functions SLOPE, INTERCEPT and STEYX
(residual standard deviation) to calculate the parameters
from the regression. The mean reversion speed is the neg-
ative of the slope, while the long run mean is the intercept

Mean Reversion Rates, Levels & Volatility Calculations

A B C D E
1 Date Current Price Price Previous
Price, $ Change, % Change, $ Price, $
2 5/1/01 15 “y” Values “x” Values
3 5/2/01 18 18.2% 3 15
4 5/3/01 15.5 -15.0% -2.5 18
5 5/4/01 12 -25.6% -3.5 15.5
6 5/5/01 14.5 18.9% 2.5 12
7 5/6/01 13 -10.9% -1.5 14.5
8 5/7/01 15 14.3% 2 13
9 5/8/01 17 12.5% 2 15
10 5/9/01 15.5 -9.2% -1.5 17
11 5/10/01 14 -10.2% -1.5 15.5
12 Standard Deviation Regression Parameters
13 STDEV(u) 16.7% SLOPE -89%
14 SQRT(250) 15.87 INTERCEPT 13.2
15 S* Annualised 265% STEYX 1.98
16
17 Long Run Mean (Intercept/Speed) 14.9
18 Volatility (STEYX/Long Run Mean) 13.2%
 Energy Pricing

estimate of that regression divided by the mean reversion

Hourly Mean Reversion Rate By Months
speed. 4.00

Mean Revision Speed (days)

6. The volatility of dollar price changes is given by the resid- 3.50
ual standard deviation calculated with STYDX. If we want to 3.00
obtain a percentage volatility, we would just need to divide 2.50

by the long run mean (see E18). 2.00

In contrast with the “random walk” process presented in 1.50

1.00
the previous issue, where price changes were independent
0.50
through time, a mean reverting process is characterised by
0.00
prices that have some degree of memory about previous

Jan
Feb

Apr

Aug

Nov
Oct
Mar

Jul
May
Jun

Sep

Dec
price changes. The mean price levels will be our best fore-
Month
cast of future price levels. In the above example, the mean
reversion level is 14.9 (Intercept value divided by the speed). 2. Mean reversion rates are not constant
The volatility of our forecast would be around 13.2% of the The speed at which prices revert to their long run levels
forecast price level, which is substantially lower than the may depend on several factors such as the nature, magni-
265% calculated with the square root of time rule. tude and direction of the price shock. If we calibrate the
mean reversion rate for each month of the year using data
Pitfalls Of Using Mean Reversion Processes To Model exclusively from that month, we would find that for most
Mean Reversion: markets the mean reversion rates differ considerably.
1. Energy Prices exhibit price jumps not described by the log In the following chart, we can see the mean reversion speed
normal distribution for hourly prices calculated for each month of the year using
Energy prices in general and electricity prices in particular, Nordpool hourly price data and FEA’s Forward Curve Builder.
diverge significantly from the log normal distribution. We can clearly see how the mean reversion rate is differ-
Although Geometric Brownian motion with mean reversion ent for each month. When using the model for valuation or
adequately models the way in which prices diffuse back risk management purposes we need to make a choice
towards the long-run equilibrium level after a jump event between using a mean reversion rate specific to each month
(eg. unexpected outage, extreme weather, etc.), it fails to or an average one extracted from the whole data sample.
capture the jump event itself.
This pitfall becomes extremely important for risk manage- Conclusion
ment and pricing of exotic and deep out-of-the-money Mean reversion in energy prices is well supported by empiri-
options, and the valuation of assets such as certain peaking cal studies of energy price behaviour, as well as by basic micro-
plants that are only turned on in extreme price scenarios. economic theory. General diffusion models that incorporate
mean reversion go a long way in capturing the nature of ener-
gy prices; notably their tendency to randomly oscillate away
from, and over time back towards a price level determined by
the cost of production. These models are gaining more wide-
spread acceptance among market practitioners as advances
UNICOM and FEA ran a successful 3-day seminar on Value-at- are made in the techniques used to estimate the mean rever-
Risk and Energy Derivatives in London during May. After an sion level and mean reversion rates. The advantages of mean
introductory day providing background to the general concepts of reversion option pricing models over their Black-Scholes coun-
VaR delegates split into two streams, one for financial markets terparts are becoming more and more apparent as traders and
and the other for energy. FEA speakers presented a wide range of risk managers are able to assign greater accuracy to their
talks on leading edge energy issues. Carlos Blanco covered models, or at a minimum to their model assumptions.
valuing assets as derivatives, Chris Brady talked about weather Unfortunately, these models fail to describe what is, for risk
derivatives and Cheryl Morgan talked about risk management in management purposes, perhaps the most important energy
the UK’s new electricity market, NETA. The seminar attracted a spot price phenomenon: the discontinuous price jump. In the
wide range of delegates including some from the largest oil, next article, we will explore the nature of these jumps and the
natural gas and electricity firms in Europe. jump diffusion price process that describes this behaviour ■

A highlight of the energy stream was the introduction of FEA’s new

Carlos Blanco is Manager of Support & Educational Services at Financial
gas storage product. King Wang of FEA explained how the product
Engineering Associates. David Soronow is a derivatives specialists at FEA.
allowed gas storage to be valued by modern contract analysis The authors would like to thank the FEA Financial Engineering & Support
methods whilst still taking into account all of the many physical and teams, especially Angelo Barbieri & Cheryl Morgan for helpful comments
practical constraints of operating a real gas storage system. and excellent software assistance. Email: [email protected]
www fea.com
Details of forthcoming seminars from Unicom
References
can be found at: 1. Financial Engineering Associates. @Energy User’s Guide. Release 2.0. 2001
2. Garman, Mark B. 1998. Managing the Risk of Fixed Assets (Jul 4-5) Derivatives
www.unicom.co.uk Strategy
3. Shimko, D. & St. Germain, J. “Substandard Deviations”, PMA OnLine Magazine,
May 1998.

June 2001 * Commodities Now 72