URLs in this article:
of Timing of Access
to the US Retail Power Market
by Anne Ku (May 1997)
Presented in May
1997 at INFORMS Spring 97 Conference in
San Diego, California
to a fully competitive retail power market depends on a myriad of factors
through a process which is largely political and highly uncertain. Individual
states vary widely in terms of the pace of decision making and issues
in restructuring. A four month study was conducted to investigate and
implement a method of answering the question: "When will we have
access to the first retail customer in each state?" Several approaches
to predicting the timing of retail access are proposed and assessed with
respect to four criteria: simple, systematic, defensible, and acceptable.
A pilot study found two of these most suitable: behavioral aggregation
of judgment-based group scenario-contingent assessment and the non-judgment
based event tracking approaches. The two approaches were implemented to
track the timing and likelihood of access on a regular basis.
Rationale for Predicting Timing of Retail Access
The deregulation and restructuring of the giant US electricity supply
industry is underway in every state. Once opened up, the US retail power
market is estimated to command $200 billion per year in revenues, which
is considerably greater than natural gas and other deregulated industries.
Whereas the deregulation of the natural gas industry took more than a
decade, the power industry is expected to go through the process in half
that time. Currently, except for pilot programs implemented in a handful
of states, customers are not allowed to choose their electricity supplier.
With retail access, customers and electricity suppliers will be able to
choose. The question of how soon they will be able to choose or "access"
depends on a political process of resolving various issues, agreeing to
a date of access, and implementing the final policy elements. The speed
of this process varies from state to state.
Enron is the largest
independent wholesale power marketer in the US. Retail, i.e. access to
the final end user, presents an enormous business opportunity. To tap
into this market, Enron requires a method for tracking the timing and
likelihood of access. This four month project was undertaken to develop
a simple, systematic and defensible
approach of predicting timing of deregulation that is acceptable
to the user of this information. A simple approach refers to one that
could be implemented without detailed involvement of many parties through
a process that is transparent. "Systematic" describes a method
that is straightforward and could be automated, without too many interventions
or exceptions. It is important that these predictions are kept up-to-date.
"Defensible" is equivalent to scientifically acceptable. User
acceptance may very well depend on communication of the methodology and
This study focuses
on predicting the timing of start of retail access only, not phase-ins
of different market segments. Although methodologies for other dates are
discussed, the results pertain to predicting the year that the first retail
customer in each state can be accessed.
The timing of
deregulation refers to the approval, start, and phase-in dates of retail
Participants in the restructuring discussions describe the process as
one of on-going debate about various controversial issues until a consensus
is reached (Raab, 1994). Experience thus far has shown it to be
a long and laborious process, one that is highly political as it involves
stakeholders with different agendas and degrees of influence. Most states
require legislative approval of retail access, i.e. mere regulatory approval
is not sufficient. Such restructuring orders typically declare start dates
of access and subsequent phase-ins. However, the actual date of implementation
depends on the resolution of various issues and the readiness of systems
and procedures put in place.
One can view the
road to retail access as passing through five stages. The first stage
is an inquiry into retail access. Most states are in this stage, except
for those which have either not started the inquiries or have decided
that retail access is not desirable. The inquiry stage is followed by
the development of a policy decision, usually at the regulatory level.
Typically policy recommendations and a timetable are taken to the state
legislature, where rule-making occurs. At each stage, certainty about
the timing of deregulation increases. With each stage, the uncertainty
about dates decreases. Table 1 shows an example of the degree of certainty
one can expect from state progress. The probability ranges reflect the
authors assessment based on interviews with various experts.
complicate this task of predicting the timing of deregulation. To investigate
suitable approaches, one needs to understand the process, identify willing
well-informed experts, and aggregate individual judgments. The task of
time prediction, especially of an outcome of a political process, is far
more difficult than predicting the outcome (winning candidate) of an election
or the verdict (guilty or not guilty) of a lawsuit as it is not just about
whether or not retail access will be approved but when it will happen.
At time of writing, there are no external entities that give such predictions
on a regular basis. Meanwhile experts who are informed and confident enough
to give such assessments proved difficult to find. Even so, their opinions
The investigation consisted of literature review, interviews with experts,
pilot studies, understanding the deregulation process, and consultation
with experts at conferences. This problem falls in the domain of judgmental
forecasting and behavioral decision making. Should we use experts or non-experts?
Should we use a judgmental or a non-judgmental (i.e. quantitative) method?
If we elicit judgments, should it be individual unaggregated, mathematically
aggregated, Delphi, or behaviorally aggregated (through a group meeting)?
What is the easiest and best elicitation method: final probabilities,
e.g. p(1); conditional probabilities, e.g. p(1|a); verbal assessment,
e.g. "most likely"; or pairwise comparison, e.g. "p(1)
> p(2)." If non-judgmental, should we use frequentist as opposed
to subjective probabilities or build a statistical model? [See Kleindorfer
et al (1993) p. 103 for a discussion of the three major schools of
probability: classical, frequentist, and subjective.] The following figure
illustrates my framework for investigating possible approaches.
An unaggregated individual assessment may be subject to bias, as described
in the seminal work of Kahneman et al (1982) on heuristics and
biases of individual judgment. Aggregation (mathematical or behavioral)
of individual judgments produces somewhat better judgments with less variability
if there is heterogeneity of opinion. A mathematical aggregation involves
either simple or weighted averaging. Ferrell (1985) notes that
an individual expert assessment is often better than an aggregation of
non-experts, as ignorance cannot be averaged out. The method of mathematical
aggregation of probabilistic judgments has strong predictable effects
on calibration and accuracy. The Delphi method refers to an iterative
method of obtaining an individual assessment, aggregating it with other
assessments, feeding back the result to the individual to get a re-assessment
until one assessment is achieved. Ferrell generalizes that Delphi refers
to any procedure for assessing opinions that uses anonymous responses,
some kind of feedback, and at least one iteration. Behavioral aggregation
refers to group consensus through negotiation.
methods involve using objective data such as whether or not an event has
occurred, the duration of an event, the actual outcome of a decision,
the number of pro-competition advocates, etc. Some combination of these
statistics will give a frequentist probability distribution. There are
also statistical modeling techniques that may be used to build a better
predictor than individual indicators.
Table 2 summarises
the main approaches investigated, with the first four methods pilot tested.
Group Scenario-Contingent Assessment
requires a group of assessors to agree on how likely access will be achieved
by a certain year. There are three scenarios: best, base, and poor. Best
case refers to the situation where we are successful in influencing the
process, the decision makers, and everything else goes in our favor. Base
case reflects the status quo.
are given for p(a), p(b), and p(c) where a = best case, b = base case,
c = poor case. To get the probability that retail access will start in
a particular year, regardless of which case (scenario), we must use Bayes
Theorem and conditional probabilities, which have not been elicited. The
probability that retail access starts in year: p(x) = p(x|a)p(a) + p(x|b)p(b)
+ p(x|c)p(c), where x = start of retail access in year 1, 2, or 3. The
conditional probabilities p(x|a), p(x|b), and p(x|c) have not been defined.
If p(1|a) = 1, p(2|b) = 1, and p(3|c) = 1, then p(1) = p(a), p(2) = p(b),
and p(3) = p(c). See figure 2 for an illustration.
encompasses a range of decision theoretic techniques, including decision
trees, influence diagrams, and multi-attribute utility analysis. Decision
trees and influence diagrams are complementary structuring tools for decision
analysis. Decision trees capture the chronological sequence of decision
and uncertain events, while influence diagrams capture conditionality
and dependence characteristics of events. Besides being more compact than
decision trees, influence diagrams reveal probabilistic dependence and
information flow. See Ku (1995) for further details. An illustrative influence
diagram in figure 3 shows that retail access dates will not be announced
unless critical issues such as stranded cost recovery are resolved.
This process of getting a probability distribution on timing of access
involves 1) interviewing well-informed experts individually, 2) drawing
an influence diagram to capture the main events, 3) drawing a corresponding
decision tree, 4) eliciting conditional and marginal probabilities, and
5) calculating them to get the final probability distribution. Figure
4 gives one way of formulating the conditioning events, in this case,
state and federal legislation only.
The decision analytic
approach appeals to the causal thinker. However, it requires the assessor
to think conditionally or a facilitator to do the elicitation. Some assessors
find it difficult to give probabilistic estimates. Over time, new events
will appear, and the decision tree will need to be restructured to include
them. For example, at time of writing, the probability that federal legislation
will be introduced is quite high, and it is expected that it will impact
state legislation. However, an unforeseen event such as a major power
outage or a competing bill may lower the probability of federal legislation,
such that another critical event should be included in the tree instead.
This is also a relatively time-consuming process as there are many possible
The above decision
tree structure could be used for scenario analysis and serve as a means
to validate the holistic assessments of the judgmental method. These decision
trees will need to be modified for those states, in particular New York,
which do not require state legislation for rule-making on retail access.
The development of generic decision tree structures for different clusters
of states will help simplify the structuring and elicitation effort.
Survey or Polling Method
The survey method
is suitable for gathering input from a large number of anonymous respondents.
In this case, respondents must be well-informed about the deregulation
issues. Although there is no face to face elicitation required in sending
out surveys, there is considerable effort in interpreting and collating
the responses. Unless the questions are made explicitly clear, there is
always a need to question the respondents assumptions. The accountability
concern mentioned earlier is removed if the respondents are anonymous.
Anonymity, however, prevents going back to the respondent for clarification.
Due to the dynamic nature of restructuring events, it is necessary to
get all the responses back quickly and simultaneously as well as to quickly
assimilate the results before they get out of date. This timeliness requirement
posed most troublesome as the respondents had no incentive to respond
other than "doing a favor."
of event-based state progress helps to validate the scenario-continent
timing assessments, gauge the extent of state progress in the process,
track the pace of deregulation, and anticipate next steps. It will also
help guide the assignment of conditional probabilities in the judgmental
method. Finally, it appears from experience so far that several issues
require resolution before retail access can occur. Such "critical"
events as stranded cost recovery, ISO formation, and divestiture (to allow
a level playing field) should be included in state monitoring.
As there are overwhelmingly
more states that are considering retail access than not, it is possible
to see patterns and trends in state progress over time. If we treat deregulation
as a process that all states have to go through, then the event tracking
method is equivalent to setting the milestones and monitoring the critical
path of a project. The process is defined by sequential "project"
stages. The period of time each state spends in each stage varies but
is usually known with some degree of certainty. By use of accelerating
and delay factors, we can track whether certain "announced"
dates will be reached or not. Similarly, by tracking how close to resolving
the critical issues, we can track how close to meeting the target date.
There are several hypothetical effects that have intuitive validity in
the event tracking approach, as follows.
Adjacent State Effect: Electricity moves beyond state borders.
Thus adjacent state activities have great influence on a states
progress. For example, New England states are somewhat synchronized
as compared to states which are not adjacent to each other.
Domino Effect: Once a good solution has been found to resolve
a controversial issue, it could very likely be considered elsewhere,
thus speeding up the process. For example, Californias rate reduction
bonds for stranded cost recovery introduces securitization as a means
to financing transition charges. Subsequently, this securitization concept
was adopted in the Pennsylvania restructuring.
Snowball Effect: Restructuring will gather momentum, especially
if certain dates need to be met such as the need to reach a decision
before the legislative session ends or a changeover of staff, especially
replacement of the Commissioner. Other related events may also accelerate
Many states in
the advanced stages of restructuring have posted timelines of expected
start date and duration of each activity. Some consultants believe there
may be delays due to the sheer number of issues that most be resolved
and the physical realities of getting systems and procedures ready.
A number of external
independent agencies monitors restructuring activity. The National Regulatory
Research Institute provides monitors the restructuring activities of each
state. Their internet web site is http://kcgl1.eng.ohio-state.edu/~hoag/nrri.html.
The bi-monthly LEAP Letter summarizes deregulation activity state by state,
as well as tracks whether or not a bill has been proposed or adopted.
method is to cluster the states by similar stages in the deregulation
process and similar predictive characteristics, such as similar circumstances
(e.g. high retail rates, liberal policies, presence of strong competition
advocate), geographic proximity (adjacent state activities), historical
bias, and proportion of Investor Owned Utilities (IOU), Cooperatives (Coops),
Municipals (Munis). Some of these characteristics have been identified
in the survey results discussed in an earlier section, but due to low
response rate, they have not been used for clustering. An example of such
clustering is the tier classification method of Regulatory Research Associates
(RRA) as shown in table 3. Another deregulation monitoring source John,
Hengerer & Esposito (JH&E) groups the states into four tiers representing
gas and power deregulation progress. Tier one is unbundling; two is emerging
states; three is potential states; four is monitor states..
A stakeholder is someone who has an interest in a decision, is affected
by the decision, and can influence the decision making process and outcome,
but does not make the decision his/herself. In economics, stakeholder
theory concerns the redefinition of property rights by providing a voice
in a firms decision making process to individuals or groups who
are affected by a firms actions, but are not necessarily shareholders
or otherwise associated with the firm by conventional means of ownership.
Stakeholder analysis involves the identification of stakeholder types,
identification of contentious issues, understanding of each stakeholders
position on each issue and negotiation strategy, and analysis of interactions.
In electricity industry restructuring, participants represent a diverse
set of interests. These stakeholders mainly fall into the following categories:
legislatures, energy service companies, retail merchants associations,
marketers, commercial business interests, utilities, ratepayer advocates,
public utility commissions, environmental groups, low-income advocates,
industrial user groups, governors offices, residential interest
groups, and independent power producers.
By assessing the relative weight of influence, we effectively produce
a decision analysis model using frequentist probabilities attached to
a stakeholders position on an issue, i.e. outcome of a decision.
The combination of a stakeholders position and relative weight of
influence can be used to assess whether there will be acceleration or
delay to meeting the date of approval and access. Aside from stake-holder
influence and interaction, there are external events or random factors
that may also affect the outcome. For example, it is well known that decision
makers, e.g. Commissioners or Senators, do not always use all information
available to make a decision, or act as one might expect. A better understanding
of behavioral decision making may be useful in this respect.
Stakeholder analysis sets the stage for other kinds of economic analyses,
such as game theory (Roth, 1985) and causal simulation (system
dynamics). These types of analyses improve our understanding of interactions
and what might happen.
The methods discussed so far rely on an absolute scale of timing in year
of access. Another approach is to rank the states relative to each other
with respect to progress. Progress can be defined by the tier classfication
system of RRA described earlier or other event tracking milestones. Although
we can similarly create key characteristics, such as critical events or
time to resolution of certain issues, as criteria for relative ranking,
they do not easily translate into probabilities.
Aside from cluster and factor analysis alluded in a previous section,
there are other statistical methods that hold potential. Assuming that
all states will go through the same process, we should be able to expect
those factors that predict for a certain state also be used for other
states. Limited methodological discussions with decision scientists indicated
that survival modeling and Bayesian regression are possible approaches.
However, these have not been investigated further in this context.
Originated in the biological and medical field, survival modeling refers
to the use of survival data to predict survival time. Survival time is
broadly defined as the time to the occurrence of a given event, e.g. death,
or in our case, approval of retail access and start date of retail access.
Survival data can include survival time, response to a given treatment,
patient characteristics, or in our case, time taken in each stage of state
progress towards deregulation, state legislatures response to public
utility commissions recommendations, states tier group, and
other state characteristics. The output of a survival model is a probability
Bayesian regression falls into a class of linear statistical models of
the form y = f(x1, x2, ..., xn). The
dependent variable y is the time to direct access in years. There is a
probability distribution around y. The independent variables are indicators,
which the same group of experts have agreed to have strongest influence
in predicting the time to access. Probability distributions are revised
over time, so the model gets refined, i.e. through posterior updates of
prior distributions. See Broemeling (1985) for further details.
The main objective of this study was to find a simple, systematic, defensible,
and acceptable methodology for predicting the timing of deregulation (start
date of retail access) for all fifty states. Unbiased assessment can be
achieved by interviewing an objective observer, i.e. one who is not a
stakeholder, and also balancing the views from representative sample of
stakeholders involved in each state. After pilot testing four out of seven
possible approaches, we recommended a parallel approach of using both
judgmental and non-judgmental methods to get monthly assessments. The
judgmental method takes scenario-contingent expert assessments from those
participants actively involved at the state and federal levels. The non-judgmental
method refers to monitoring state progress by use of internal indicators
and external monitoring entities whose results are published regularly.
Due to the timeliness of the results, it is necessary to ensure the monitoring
sources are in sync with each other as well as with the expert assessments.
The difficulties encountered in most of the approaches are not insurmountable
as shown in table 4. They can largely be overcome by management directive
and sufficient resources. Alternatively, they may be removed altogether
if the exercise is conducted by an external entity who not only monitors
state progress but can also ensure full response rate from a regular (repeatable)
survey of a representative sample of stakeholders in each state and collated
and updated in a timely manner.
As the pace of restructuring accelerates and the learning curve flattens,
we can expect to find more experts. Existing experts will be able to see
patterns in state progress to predict timing more accurately and with
greater confidence. Similarly there may be more entities reporting and
monitoring state progress, providing further "validation" to
Figures and Tables
Table 1 Timing Certainty (illustrative probabilities
that the year will be known, discussed, or will occur as said)
Figure 1 Possible Approaches
2 Comparison of Alternative Approaches
Figure 2 Conditional Probabilities
Figure 3 Influence Diagram
Figure 5 Probability Distribution
Table 3 RRAs Classification
Table 4 Difficulties and Suggestions to Overcome