Five Ways to Ensure that Models Serve Their Purpose

 Published on July 12, 2020

Burney Waring
Director of Retirement Testing at Waring Retirement Laboratory

The authors of this manifesto in Nature wrote five points to consider about epidemiological modeling and
decision-making for society. These are worthy reminders to modelers in all fields, including engineers
working in the oil and gas industry. I am quoting extensively from their article here, which is worth reading
in full.

“We are calling not for an end to quantification, nor for apolitical models, but for full and frank disclosure.
Following these five points will help to preserve mathematical modelling as a valuable tool. Each
contributes to the overarching goal of billboarding the strengths and limits of model outputs. Ignore the
five, and model predictions become Trojan horses for unstated interests and values. Model responsibly.”

Below the five areas for an engineer to mind are their: Assumptions, Hubris, Framing, Consequences,
and Unknowns.

1. Mind the Assumptions

The first assumption the modeler should encounter is the assumed purpose for the model. No matter
what, the first purpose might not be the last purpose, so the engineer needs to keep a constant check on
the demands on the model and try hard to be insensitive to the time and effort sunk into a model that is
no longer fit-for-purpose.

“Models are often imported from other applications, ignoring how assumptions that are reasonable in one
situation can become nonsensical in another. Models that work for civil nuclear risk might not adequately
assess seismic risk. Another lapse occurs when models require input values for which there is no reliable
information. For example, there is a model used in the United Kingdom to guide transport policy that
depends on a guess for how many passengers will travel in each car three decades from now.”

Engineers, just how uncertain is your data? Start by looking at the data suspiciously. How dependent is
your decision on your assumptions about the data? Would it be profitable to get more/better data? If you
are wrong, are there non-financial losses that should be taken into account? Is the technique being used
one that is reliable in this particular situation? Should you run an uncertainty analysis, for example a
Monte-Carlo series of simulations to understand the range of uncertainty in your ultimate decision? If you
are convinced that a decision is the correct one, you should still provide the uncertainty range along with
your recommendation. You may not have the full picture of the decision-maker. Or, your honest range
may prompt other investigations.

2. Mind the Hubris

“One extreme example of excess complexity is a model used by the US Department of Energy to
evaluate risk in disposing of radioactive waste at the Yucca Mountain repository. Called the total system
performance assessment, it comprised 286 sub-models with thousands of parameters. Regulators tasked
it with predicting “one million years” of safety. Yet a single key variable — the time needed for water to
percolate down to the underground repository level — was uncertain by three orders of magnitude,
rendering the size of the model irrelevant.”

Complexity isn’t accuracy. Instead, complexity in a model can degrade the creator’s ability to understand
how a model yields an answer, to test the model thoroughly, or to troubleshoot the model when it seems
wrong. “As more parameters are added, the uncertainty builds up (the uncertainty cascade effect), and
the error could increase to the point at which predictions become useless.”

Beware of averages. The average reservoir behavior in an unconventional reservoir is irrelevant when
production is coming from a few areas where fracs are productive. The average person in a neighborhood
is irrelevant if most of the spread of a virus is caused by a few super-spreaders.

“Complexity is too often seen as an end in itself. Instead, the goal must be finding the optimum balance
with error. What’s more, people trained in building models are often not drilled or incentivized for such
analyses. Whereas an engineer is called to task if a bridge falls, other models tend to be developed with
large teams and use such complex feedback loops that no one can be held accountable if the predictions
are catastrophically wrong.”

Someone (preferably the whole core team) must keep the overview. If they cannot, the thing is too
complex to be managed.

3. Mind the Framing

“Match purpose and context. Results from models will at least partly reflect the interests, disciplinary
orientations and biases of the developers. No one model can serve all purposes.”

A production system model used for optimizing development scenarios is useless for production
optimization. A reservoir model is useless for forecasting if it cannot predict limitations in wells and
facilities. If a model can incorporate on/off behaviors of wells and equipment, but cannot predict reservoir
behavior it might be perfectly useful for a 90 day forecast, but useless for a 1 year forecast.

“Modellers know that the choice of tools will influence, and could even determine, the outcome of the
analysis, so the technique is never neutral. For example, the GENESIS model of shoreline erosion was
used by the US Army Corps of Engineers to support cost–benefit assessments for beach preservation
projects. The cost–benefit model could not predict realistically the mechanisms of beach erosion by
waves or the effectiveness of beach replenishment by human intervention. It could be easily manipulated
to boost evidence that certain coastal-engineering projects would be beneficial. A fairer assessment
would have considered how extreme storm events dominate in erosion processes.”

I know engineers that ‘massage’ data to produce the outcomes desired by their managers. This is a
losing game for an organization. And, for an individual, it can be a lose-lose bargain with such
managers. Your reputation and your sanity are important things to consider. Don’t give into such requests
easily. Usually if you play ignorant and ask for things in writing, these requests go away.

“The best way to keep models from hiding their assumptions, including political leanings, is a set of social
norms. These should cover how to produce a model, assess its uncertainty and communicate the
results.”

Within companies, there should be standard processes which expose the techniques used and data to
reviewers of model results. In any kind of data analysis situation, the code for the analysis can be
published with the results. If there isn’t such a process, create and use one and document it for
others. They will probably thank you.

4. Mind the Consequences

What are the consequences of your recommendation? Is there another solution that will on average
provide worse economics, but is better in the less likely but more problematic ranges of your data? (I.e. a
safer option, if things go wrong.)

“Quantification can backfire. Excessive regard for producing numbers can push a discipline away from
being roughly right towards being precisely wrong. Undiscriminating use of statistical tests can substitute
for sound judgement. By helping to make risky financial products seem safe, models contributed to
derailing the global economy in 2007–08.”

It will be a hard fight for you, just as it has been for your predecessors, but you must continuously provide
ranges for your answers even when a single number is demanded. A rich conversation is the only one
worth having. So, when you absolutely must select a single number, you must first know how it will be
used and the consequences of that number being incorrect on the high and low sides. For example, if
you predict a project will be complete somewhere between July 1 and August 15, and you are asked to
give a single completion date, it would be useful to know that if the project is early there is no
consequence but if it is late there is a barge standing by for $500,000 per day.

“Spurious precision adds to a false sense of certainty. If modellers tell the United Kingdom it will see
510,000 deaths if no steps are taken to mitigate the pandemic, some might imagine a confidence of two
significant digits. Instead, even the limited uncertainty analysis run by the modellers — based on just one
parameter — reveals a range of 410,000–550,000 deaths. Similarly, the World Health Organization
predicts up to 190,000 deaths for Africa (see go.nature.com/3hdy8kn). That number corresponds to a
speculative scenario in which ten uncertain input probabilities are increased by an arbitrary 10% — as if
they were truly equally uncertain — with no theoretical or empirical basis for such a choice. Although
thought experiments are useful, they should not be treated as predictions.”

My first engineering course included a huge amount of tedious problems involving significant digits and
the correct precision. I wish I had paid more attention. There is an inherent promise built into providing a
number. Are you actually predicting 96,000 barrels per day? Or, are you predicting somewhere between
50,000 and 110,000 with the most likely case being about 95,000? You owe everyone the full information,
even if they don’t ask for it.
“Opacity about uncertainty damages trust. A message from the field of sociology of quantification is that
trust is essential for numbers to be useful. Full explanations are crucial.”

5. Mind the Unknowns

“Models can hide ignorance.” First, do not let your own model hide your own ignorance from
yourself. Understand your model thoroughly. You should have an intuitive, accurate feel for the model
before you can drive it safely. Second, if you are asked to predict something impossible, (“when will this
well have water breakthrough?”) acknowledge the reality and explain why (“we don’t even know in reality
if this is a bottom- or edge-water drive reservoir and do not know for sure where our injected water is
going”).

Learn to be good at explaining what you don’t know and why there are limits to your predictions. There is
a certain art to this. “I don’t know, but I can find out” is a good start. “The range on this answer is very
wide. If it is important to know more precisely, there are several things we need to do first” is better.

“Failure to acknowledge this can artificially limit the policy options and open the door to undesired
surprises. Take, for instance, those that befell the heads of governments when the economists in charge
admitted that their models — by design — could not predict the last recession. Worse, neglecting
uncertainties could offer politicians the chance to abdicate accountability. Experts should have the
courage to respond that “there is no number-answer to your question”, as US government epidemiologist
Anthony Fauci did when probed by a politician.”

Conclusion
“Mathematical models are a great way to explore questions. They are also a dangerous way to assert
answers. Asking models for certainty or consensus is more a sign of the difficulties in making
controversial decisions than it is a solution, and can invite ritualistic use of quantification.”

Engineers should not hide behind their models. Models can support recommendations but cannot be
creative, nor understand implications, nor consider trade-offs.

This is an important aspect of engineering that isn't discussed enough. I would love to hear your
comments.

Burney Waring is an almost-completely retired global consultant engineer, and Director of Retirement
Testing at the Waring Retirement Laboratory.