WEATHER RESEARCH FOR
TRADING PROFITS
© Copyright 2007 by Michael A. S. Guth. All Rights Reserved. No portion of this site, including the contents of this web page may be copied, retransmitted, reposted, duplicated, or otherwise used without the express written permission of Dr. Michael Guth. Reprinted from The Risk Desk (May 2002) with permission of the publisher, Scudder Publishing Group, LLC. www.scudderpublishing.com.
|
|
MICHAEL A. S. GUTH, Ph.D., J.D.
|
Financial Economics Homepage || Attorney at Law Homepage
Weather Research for Trading Profits
Michael
A. S. Guth
Managing Director, Risk Management Consulting
Gary
M. Lackmann
Professor, Department of Marine, Earth, and Atmospheric Sciences,
North Carolina State University
Scott
E. Kennedy
Department of Marine, Earth, and Atmospheric Sciences, North
Carolina State University
K.
Wyat Appel
Department of Marine, Earth, and Atmospheric Sciences, North
Carolina State University
Ask any power or natural gas
trader what is the single most important determinant of electricity prices, and
you will get the same answer: weather.
Weather also has a significant impact on the price of natural gas. Today,
traders have access over the Internet to more than two dozen web pages with
weather information. Traders feel inundated and somewhat overwhelmed by the
sheer volume of weather forecasts and data, and they generally feel that
sponsoring additional weather research would have no value. Additional weather research would just add
one more forecast to the pile.
It turns out these traders
are wrong. Some weather research can
indeed improve upon the accuracy of the National Weather Service and its
private-sector competitors. To get the
“good stuff,” companies generally will have to pay for the service from a
professional meteorologist. Because all
traders have access to the Internet, the publicly available information on
weather is already reflected in current prices of electricity and natural gas
forward contracts. If traders want
weather information that will give them a competitive edge, then they need to
acquire weather information that is both different
and better than what is publicly
available.
This article illustrates one
type of weather research project that power and natural gas trading firms
should want to sponsor, particularly if they have aspirations about trading
weather derivatives or making big profits from a summertime heat wave. The goal of this research project is to
produce 10-day forecasts for extreme summertime heat that are more accurate
(because they rely on different information) than the forecasts of the National
Weather Service and other public information on the Internet. If traders can score big in the summertime
months, when power prices are most volatile, they will usually guarantee that
their firms will have a profitable year on the trading floor.
Motive
Acting as independent
consultants for a power company during the summer of 2000, members of the NCSU
weather forecasting team gained knowledge on the scientific techniques for
extended-range maximum temperature forecasting. Initially, this study had an objective of providing confidence
levels for the weather forecasts. But,
as is often the case with basic research, to translate those findings into
useful information for traders required more than general statements about
confidence levels. Traders wanted
actual forecasts for the cities where they held open positions.
In their first attempt at bat
in the summer of 2000, the NCSU team’s forecasts came in slightly more accurate
(averaged over eight cities and an entire summer of 1-8 day lead-time
forecasts) than those from a widely used commercial forecasting service. However, with hindsight, the team found a
way to improve the accuracy of its forecasts and also discovered “interesting”
patterns in the forecast errors. These
patterns could be translated into lucrative trading strategies if they were
refined in the right manner.
With time series data, we
expect the usual error term problems of serial correlation and
heteroscedasticity (from missing explanatory variables). Instead of these predictable error term
patterns, the forecasts exhibited periodic “large error events,” during which
the actual temperatures were dramatically hotter (or colder) than had been
forecasted. The widespread summertime
heat events had common denominators, and these events highlighted systematic
biases in commonly used forecasting tools during certain synoptic flow
configurations. If these biases could
be mitigated in some way, then the 1-8 day forecasts of extreme heat could be
substantially improved. Clearly, traders are less interested in reducing errors
in the forecasted high temperature by one or two degrees each day than they are
in knowing when the big electricity price blowout events will occur.
Why are Extreme Heat Predictions Often Poor?
In order to gain an
appreciation for the research strategy outlined below, it is first worthwhile
to examine the reasons why many run-of-the-mill forecasts for summer heat waves
are unreliable. Firstly, the accuracy
of weather forecasts for a given type of weather event depends heavily on the
amount of forecaster experience with respect to that phenomenon. Roebber and Bosart (1996) analyzed a huge
volume of forecast information and concluded that “forecast skill is largely
determined by experience”. Because
extreme heat is a relatively rare event, professional weather forecasters have
precious little experience to draw upon in predicting these events.
Second, the statistical
algorithms used by meteorologists to generate so-called model-output statistics
(MOS, e.g., Jacks et al. 1990) are by their very nature less reliable for rare
events. MOS is based on a multiple
linear regression between past model output values and what really ended up
happening. For rare or extreme events,
the MOS regression equations are drawing upon regions of phase space that are
sparsely populated, driving the uncertainty skyward. In cases where heat reaches record-setting levels, the MOS
equations may even be forging into uninhabited reaches of parameter space.
Additionally, the MOS equations factor in climatological averages, which will
tend to pull the forecasts away from all-out predictions of extreme conditions.
Similar weaknesses are implicit in the “super-ensemble” forecasts, of the type
described by Krishnamurti et al. (2000).
Although these super-ensemble techniques are useful for extracting skill
from a variety of forecast sources, they still lose accuracy for rare events
for the same reasons that traditional MOS products do.
Finally, there are several
subtle physical processes that come to play in extreme heat events, such as
soil-moisture feedbacks (e.g., Namias 1991). Some of the fundamental
assumptions that are built in to the computerized weather prediction models
that human forecasters rely upon are violated in extreme heat conditions. The
bottom line: meteorologists must roll up their sleeves and undertake rigorous
analyses of the physics of heat events if they hope to improve predictions.
Our Research Agenda
The first phase of our
research project involved building a long-term (30-year) climatological
database of extreme heat events. Using a weighted combination of maximum and
minimum temperature, along with a measure of humidity, we developed a
meteorological parameter (dubbed the extreme-heat index, or EHI) that is more
highly correlated with energy load or price than more traditional parameters
such as daily maximum temperature, cooling-degree days, or heat index. After
removing weekends and holidays, the coefficients of determination (R2)
between the EHI, the load, and price were computed. This methodology explained nearly three-quarters of the variance
in summertime electricity load for one region in the southeastern United States
over a recent five-year period.[1] For price data (obtained for various points
in the Eastern Interconnect, including the power trading hubs at PJM, Cintergy,
Entergy, and TVA), strong correlations were again evident. For example using PJM data, this methodology
was shown to explain sixty percent of
the variance of electricity prices when the EHI exceeded a threshold value. The
point of this exercise was merely to demonstrate, using very simple statistics,
that the heat events studied are directly relevant to the power industry. Using
the EHI, we then retrospectively interrogated a 30-year database of hourly and
daily meteorological data using a statistically defined EHI threshold. The threshold was set in order to isolate
heat events that sent electric load far above average, and 44 independent
widespread heat events were thus identified.
During some summers, there
were over two week’s worth of extreme heat days. Other summers had none. Intriguing patterns (and cause-effect
relations) became apparent with respect to interannual variability. A
by-product of this analysis is statistical information regarding inter-annual
variability, including relative frequency of widespread heat as a function of
El Niño/Southern Oscillation (ENSO).
Although not a primary objective, it is possible that seasonal
predictive skill could be derived from a more detailed analysis of these data.
The power industry is most
interested in widespread heat events,
i.e., when there is high demand and short supply of electricity all across the
Eastern Interconnect. Few studies are available in the scientific
literature to provide meteorological documentation for such extreme heat
events. Given the tremendous socioeconomic impact of these events, it is
perhaps surprising that they have not received more attention. For example, a
useful forecasting parameter for maximum temperature is the 850-hPa
(hecto-Pascal, also known as a millibar (mb)) temperature: the temperature
approximately 1.5 km above sea level.
Computer model forecasts of 850-hPa temperature are typically more
accurate than the forecasts of surface temperature, because the latter are
affected by often-unrealistic model representation of surface characteristics
(e.g., land use, terrain, and soil moisture).
By using the more reliable 850-hPa temperature, forecasters can produce
accurate maximum temperature forecasts using common thermodynamic techniques. Without undertaking a complete historical
analysis, forecasters do not have sufficient knowledge of the statistical
relation between a given 850-hPa temperature and maximum surface temperature in
a given geographical region. Now that
this information is in hand, a forecasting benefit during extreme heat events
can be realized.
Composite mean fields were
constructed for various meteorological variables for lagged composite times
ranging from one week prior to the onset of the heat to three days after. Composites of sea level pressure, 850-hPa geopotential
height and temperature and 500-hPa geopotential height were chosen as the
initial parameters. In order to
identify and elucidate precursor signals, height anomalies (deviations from climatology) were computed, along with
statistical significance via a 2-sided student-t
test (see Lackmann et al. 1996 for an example of this methodology). The use of
lagged composites has allowed us to isolate subtle precursor signals in the
antecedent atmospheric fields.
Additionally, we have the ability to stratify the case sample in order
to detect differences in the patterns accompanying long-lived versus short heat
events. Other composites can be
generated for the demise of a heat wave, which will aid prediction of heat-ending
cooling trends.
The composites don't tell the
whole story. “Smearing” of atmospheric fields reduces the level of detail to
that found only in larger spatial scales and may obscure important
case-particular details. Therefore, careful climatological analysis will also
be conducted to account for case-to-case variability. Although we have developed several useful results thus far, much
work remains to be done. Case-study
analyses of past forecast failures, a forecast-error climatology, and model
weighting specifically appropriate for heat events are necessary in order to
fully optimize heat forecasting on 1-10 day time scales.
How Can
Traders Best Utilize Weather Forecasts?
The evidence suggests that
the provision of 1-10 day forecasts of
EHI, rather than maximum temperature or cooling-degree days, would be quite
helpful to traders. In principle, it is
possible that given a more rigorous statistical analysis of EHI and price time
series, actual predictions of energy price, or at least that part of the price
is explained by fluctuations in the EHI, could be generated. Although complications relating to changes
in demand and capacity would need to be accounted for, this is one way that the
forecast output could be related more directly to the information traders are
seeking. For the time being, traders
would best benefit from direct consultation with professional
meteorologists. Some additional
information, above and beyond traditional weather forecasts, could be provided
to facilitate decision processes.
To summarize, the forecasting
research team from NC State University has developed a meteorological parameter
(the EHI) that is more highly correlated with energy load and price than
traditional forecast parameters. Using
an EHI-based algorithm to interrogate historical meteorological databases
(including the past 30 years), the NC State team was able to identify a large
number of past events which bore a strong meteorological resemblance to recent
events that reflected strongly on the energy market. Composites were then
produced and analyzed in a manner which will allow future cases to be predicted
with much greater accuracy and confidence. A cursory glance at the composites
has revealed some interesting and unanticipated meteorological signals that
will be advantageous in forecasting and trading profits. To date, these preliminary results have
exceeded even the research team's optimistic expectations.
References
Krishnamurti,
T. N., C. M. Kishtawal, Z. Zhang, T. LaRow, D. Bachiochi, and E. Williford,
2000: Multimodel ensemble forecasts for
weather and seasonal climate. J. Climate 13, 4196–4216.
Jacks,
E., J. B. Bower, V. J. Dagostaro, J. P. Dallavalle, M.C. Erickson, and J. C.
Su, 1990; New NGM-based MOS guidance
for maximum/minimum temperature, probability of precipitation, cloud amount,
and surface wind. Wea. Forecasting, 5, 128–138.
Lackmann,
G. M., L. F. Bosart, and D. Keyser, 1996:
Planetary- and synoptic-scale characteristics of explosive wintertime
cyclogenesis over the western North Atlantic Ocean. Mon. Wea. Rev., 124, 2672–2702.
Namias,
J., 1991: Spring and summer 1988
drought over the contiguous United States—Causes and prediction. J.
Climate, 4, 54–65.
Roebber,
P. J., and L. F. Bosart, 1996: The
contributions of education and experience to forecast skill. Wea. Forecasting, 11, 21–40.
Professor
Gary M. Lackmann can be reached via e-mail at gary@ncsu.edu, and by phone at
(919) 515-1439. His web page is http://www4.ncsu.edu/~gary/forecastlab/ Dr. Michael Guth can be reached at
e-mail mike
at michaelguth.com and phone (865) 483-8309.
Graduate students Scott Kennedy and Wyat Appel can be reached at e-mail sekenne2@unity.ncsu.edu and
wkappel@unity.ncsu.edu.
This
research was initially pursued due to the visionary leadership of Greg Locke,
Section Head in charge of forward and spot power trading at Progress Energy
from 2001 - 2002. Locke can now be reached
at e-mail address glocke@electricities.org
©
Copyright 2007 by Michael A. S. Guth. All Rights Reserved. No portion of this
site, including the contents of this web page may be copied, retransmitted,
reposted, duplicated, or otherwise used without the express written permission
of Dr. Michael Guth. Reprinted from The Risk Desk (May 2002) with
permission of the publisher, Scudder Publishing Group, LLC. www.scudderpublishing.com.