Professional Documents
Culture Documents
Articles by Forecasters
Articles by Forecasters
ARTICLES by FORECASTERS
for FORECASTERS: issue 63
Into the (Largely) Unknown: A Simple Way to
Handle Uncertainty
PREVIEW In this first of a two-part article, Steve Morlidge and Paul Goodwin make a strong
case for the use of possibility distributions to estimate likelihoods of different outcomes (e.g.,
sales levels) in situations, such as new-product introductions, where there are no past data to
provide guidance.
When we face decisions relating to an uncertain future, it’s best not to pretend that
we can make precise forecasts about what will transpire: there’s no value in being
precise about our ignorance. In these cases, simple, transparent methods that formally
recognise our lack of knowledge, and the fuzziness that accompanies it, come into
their own.
14 FORESIGHT 2021: Q4
for example whenever we launch an in-
novative product or enter a new territory. Key Points
In scenarios like those described by
■ In many decisions, such as whether to launch
Scoblic and Tetlock (2021) in the previ-
ous issue of Foresight, there is another a revolutionary new product or to invest in
factor that makes prediction much more a potentially groundbreaking technology,
difficult. Forecasts made by an electricity we have no relevant past data and little
company are unlikely to change people’s knowledge to guide us on what future
behaviour – most people will be unaware outcomes might be. The lack of a reference
of them. But some forecasts do have this class of past similar events makes it difficult to
power, and this quality of reflexivity can estimate probabilities for these outcomes.
mean that a forecast will be wrong as
soon as it is made. ■ In these circumstances, possibility theory offers
a simple and flexible alternative approach
In several elections in the UK, it is thought that enables us to make quantitatively based
that opinion-poll-based forecasts of a
statements, despite the lack of precision
clear victory by one political party caused
many of its supporters to stay at home on inherent in our estimates of the future.
polling day, thinking their vote was not ■P
ossibility theory is based on the idea of
needed. The result: a less clear-cut vic- fuzzy sets, which allow imprecise or vague
tory than the one forecast (Norris, 2001) information to be represented in mathematical
or even a surprise victory for the other terms and hence treated formally. Fuzzy-
party. Medium-term business forecasts set theory and possibility theory have been
are explicitly reflexive because they are
applied in a wide range of fields, including
used to steer the performance of a busi-
ness. For example, a firm can take action facial recognition, control of subway systems,
to try to ensure that a projected decline in medical diagnosis, database querying, and
turnover does not come about. fault diagnosis in satellites.
■ In this and a follow-up article, we demonstrate
RISK AND UNCERTAINTY
how the approach can be applied to a range
Frank Knight, the University of Chicago of decisions when the future we face is highly
economist, in 1921 drew a sharp distinc- uncertain.
tion between circumstances where we can
measure the chances of events occurring
and those where we lack sufficient knowl- In these situations of uncertainty, we
edge to allow us to quantify the prospect cannot use the probability tools of con-
of such events. Knight argued that in the ventional forecasting. When there is no
first situation, where we can use prob- reference data upon which to base calcu-
abilities, we face risk. In the second, we lations or judgments, probabilities do not
face a fundamentally different challenge: have any meaning. But does that mean
uncertainty. that, in conditions of uncertainty, there is
In cases like this, we may have to admit no place for calculation or rationality? If
that we have little or no idea what might that were the case, human beings would
happen, either because we could know be in trouble. We believe that there is a
but we (currently) don’t or because the solution, and it involves building on the
future is fundamentally unknowable. heuristics that we use in everyday life to
Much of the future will be affected by deal with uncertainty.
developments in technology, geopolitics, We start with the fact that in life we are
economic shocks, and situations where rarely totally ignorant about the pros-
human agency plays a major role (e.g., the pects for the future. We usually know
reactions of a competitor). These are all enough to construct a plausible narrative
areas where it may be wise to make such (or set of narratives) about what might
an admission. happen or be the case. And intuitively we
https://foresight.forecasters.org FORESIGHT 15
know when a new fact is consistent with because we have no reference class to
this narrative. When it isn’t, we are sur- inform our judgments, and as a result we
prised or shocked and talk about some- may be swayed by factors such as wishful
thing “straining the bounds of credibil- thinking or political bias.
ity.” And as time goes by and we acquire But if it were possible to specify our be-
knowledge, the central narrative should liefs in a way that enabled us to quantify
change and the bounds of what is conceiv- judgements about plausible outcomes, it
able will narrow. If this sounds familiar, could enable us to test and refine them,
it is because it isn’t too different from a have informed discussions, and make ra-
process that we are very familiar with: tional decisions – even in the absence of
the scientific method. This always starts definitive knowledge.
off with a judgement called a hypothesis
which is subsequently tested and refined POSSIBILITY THEORY
empirically. This suggests that it should
be possible to operationalise and quantify In situations like this, where we have some
this approach to dealing with uncertainty, knowledge but not enough to estimate
starting with a narrative that is probabilities, possibility theory can offer
• Plausible a way forward (Dubois and Priad, 2011).
Possibility theory is based on the idea of
• Internally consistent fuzzy sets, which allow imprecise or vague
• Capable of being empirically tested or information to be represented in mathe-
falsified. matical terms and hence treated formally
However, we need to be careful here. (Zadeh, 1999). We often naturally think
People are brilliant at inventing narratives in terms of concepts that are ill-defined.
to explain disparate events. This is espe- For example, we may describe someone as
cially the case when the narrative is self- “young” or the weather as “warm.” Both
serving or in accord with the way we want the set of ages considered to be “young”
to see the world. And when disconfirming and the set of temperatures that might be
facts arrive, we tend to resist changing our considered “warm” have fuzzy boundar-
views, selectively seeking evidence that ies. Is 61 degrees Fahrenheit warm, or 68
favours what we believe. Scientists often degrees?
hold on to their preferred theories even Fuzzy Sets
as strong evidence against them mounts, In these cases, fuzzy-set theory allows
only changing their minds when the case us to assign a number to represent the
against them is overwhelming (Kuhn, plausibility that a given temperature con-
1962). Approaches that encourage people forms to the term “warm.” If we do this
actively to seek, and take account of, dis- for all temperatures, we have a possibil-
confirming evidence – testing narratives ity distribution. Like probability theory,
to destruction – can act as an antidote to possibility theory uses a scale ranging
this tendency. In forecasting teams, such from 0 to 1, where a value 0 is ascribed to
devil’s advocates can play a valuable role potential states that are considered to be
in challenging narratives that otherwise highly implausible in the context of the
might be accepted without question. current narrative, while 1 represents out-
The good news is that, even when pros- comes that are the most credible or would
pects are uncertain, many business prob- be totally unsurprising if they occurred.
lems at least lend themselves to precise Suppose a friend phones to tell you that,
questions, such as how much of this new where she is, the weather is “warm.” This
product will we sell or what net present value could suggest the following possibilities:
(NPV) will this business project generate? 0 for 35 degrees Fahrenheit, 0.5 for 55
The challenge that we are currently faced degrees, 1 for 75 degrees, 0.75 for 85
with is that, if the new product or proj- degrees, and 0 for 115 degrees. (We’ll dis-
ect is unique, we cannot credibly answer cuss later how numbers like these can be
these questions with exact probabilities derived.)
16 FORESIGHT 2021: Q4
Possibilities vs. Probabilities Figure 1. A Possibility Distribution for a New Product’s First-Year Sales
Possibility theory offers greater simplicity
than probability theory. Unlike probabil-
ity theory, for example, the possibilities
for a set of mutually exclusive and exhaus-
tive outcomes are not constrained to sum
to one. Indeed, several alternative out-
comes can each be assigned the maximum
possibility of 1. Moreover, when a new
plausible future event is identified, pos-
sibility theory recognises that this does
not necessarily reduce the plausibility of
other possible events, unlike the case with
probability theory. Nevertheless, while
possibility distributions are more ap-
propriate than probability distributions while the most plausible level of sales is
when our information is imprecise, there judged to be 8,000 units.
are analogies and some weak connections
between the two concepts (Zadeh, 1999). This simple triangular shape makes it
arithmetically easy to calculate areas
Fuzzy-set theory and possibility theory within the triangle and thus quantify
have been applied in a wide range of fields, possible ranges of outcomes, such as sales
including facial recognition, control of exceeding particular levels. Calculating
subway systems, medical diagnosis, data- a single value to represent the distribu-
base querying, and fault diagnosis in sat- tion (its so-called centre of gravity) is also
ellites. And it turns out that the ideas we straightforward, even when the distribu-
used to treat a vague term such as “warm” tion is skewed. Crucially, the triangular
can also be applied when we have to make shape makes no assumptions about the
decisions relating to an uncertain future (uncertain) nature of the distribution
– for example, in areas like investment of possibilities around the most plau-
appraisal. So these ideas are not entirely sible outcome, and so limits the degree
new: over 60 years ago, the Liverpool of speculative judgements that are taken
University economist G.L.S. Shackle into account. Note that distributions with
(1961) formulated surprise theory, which other shapes are also used in practice. For
involves estimating values that reflect the example, a uniform shape would be ap-
extent to which you would be surprised propriate where all values between the
by a given outcome if it occurred. minimum and maximum are considered
The rest of this article describes how pos- to have possibilities of 1. Alternatively,
sibility theory can be applied to business a range of middle values may all have the
decisions when we are facing uncertainty maximum possibility, leading to a distri-
about the future. In a subsequent article, bution with a trapezoidal shape (Shaheen
we will discuss a range of real-world ap- and colleagues, 2007).
plications for this approach. Suppose that sales of 9,000 units are
needed for the product to break even. The
APPLYING POSSIBILITY DISTRIBUTIONS
possibility distribution can be used to de-
A commonly used possibility distribution termine a success score of turning a profit in
has a triangular shape like that shown in its first year. This possibility is represented
Figure 1 (Borges and colleagues, 2018), by the shaded proportion of the area un-
which displays possibilities for the fuzzy der the triangle in Figure 1. To make this
set of what a new product might achieve calculation, we need first to determine
in sales during its first year after introduc- the area of the whole triangle: since a tri-
tion. The bounds of the set are 6,000 and angle’s area is half its base times its height
20,000 units, each having zero possibility, (the height being 1), the triangle’s area is
½ (20,000-6,000) = 7,000.
https://foresight.forecasters.org FORESIGHT 17
aa triangle’s
triangle’s area
area isis half
half its
its base
base times
times its
its height
height (the
(the height
height being
being 1),
1), the
the triangle’s
triangle’s area
area isis ½
½
(Most plausible sales – Lowest sales)
(20,000-6,000)
(20,000-6,000) == 7,000.
7,000.
A representative value of the fuzzy set of possible sales can be found by calculating the
One
One way
way to
to calculate
calculate the
the shaded
shaded area
area isis way to calculate the shaded area is
One possible
distribution’s centre of gravitysales
(CoG),can beis found
which bytocalculat-
analogous the mean of a probability
distribution. ing the distribution’s centre of CoG
Because we have a triangular distribution, the gravity
can be calculated using th
0.5
0.5 (Highest
(Highest sales
sales –– Breakeven
Breakeven sales)
sales) == 0.5(20000-9000)
22
0.5(20000-9000) == 5042
22
5042
simple formula: (CoG), which is analogous to the mean
(Highest
(Highest sales
sales –– Most
Most plausible
plausible sales)
sales) (20000-8000)
(20000-8000)
of a probability distribution. Because we
CoG = (Lowest sales + Most possible sales + Highest sales)/3
See
See the
the Appendix
Appendix for
for the
the logic
logic behind
behind this
this formula.
formula.
CoG =CoG
(6000 + 8000
= (6000 + 20000)/3
+ 8000 = 11333
+ 20000)/3 units.units.
= 11333
18 FORESIGHT 2021: Q4
We will show how the CoG can be used to obtain a ballpark figure for the value of an investment in
We will show how the CoG can be used to obtain a ballpark figure for the value of an investment in
are facing, something that is often under- No method is perfect. While the simplic-
estimated (Russo and Schoemaker, 1992). ity of possibility theory may lead to the
Nevertheless, there is a danger that the reduction in cognitive biases when com-
chosen range of possibilities will be too pared to probability theory, it is unlikely
narrow. Especially if we first establish the to eliminate these biases. In most applica-
most plausible level of sales, there may be tions of judgmental forecasting, political
a tendency to anchor on this value when gaming can rear its head and distort the
specifying the upper and lower bounds of estimation process. Moreover, while pos-
sales, creating too narrow a base for the sibility theory may reduce the tendency
triangle. to underestimate uncertainty, it could in
It is therefore important to ask deci- some cases exaggerate uncertainty. When
sion makers to reflect on reasons why two possibility distributions with a simi-
outcomes may lie outside the suggested lar shape are being added or subtracted,
ranges of possibilities, and encourage the resulting distribution has the same
them to widen the range before applying shape, but it is wider than that of an anal-
the method. Starting with ranges, rather ogous probability distribution (Shaheen
and colleagues, 2007). For example, if we
It is therefore important to ask decision makers to reflect on reasons why outcomes
may lie outside the suggested ranges of possibilities.
than most-plausible outcomes, or using a
range of scenarios to envisage the chains subtract a distribution of costs from a
of events that might lead to alterna- revenue distribution, the possibility dis-
tive future events may be helpful here. tribution for profit may overestimate the
Scenarios allow the rationale underlying uncertainty we face.
estimates to be documented and allow as- CONCLUSION
sumptions to be made explicit, and hence
open to challenge. When we face decisions relating to an
uncertain future, it’s best not to pretend
Remember that the possibility distri- that we can make precise forecasts about
bution reflects what people think will what will transpire: there’s no value in be-
happen, and minds will change as new ing precise about our ignorance. In these
knowledge is acquired – a Bayesian-style cases, simple, transparent methods that
updating process. Indeed, the possibil- formally recognise our lack of knowledge,
ity distribution can be used to guide the and the fuzziness that accompanies it,
acquisition of new knowledge that can come into their own. By making people’s
reduce the spread from worst to best judgments explicit, acknowledging un-
outcome and especially reduce the loss- certainty, and emphasising the need to
making part of this range. test and, if necessary, modify beliefs by
Finally, possibility theory allows us to acquiring new evidence, the method we
assess how well our possibility estimates have described should lead to more real-
performed. An outcome that occurs istic forecasts. In Part Two of this article,
despite given a low or zero possibility we will explore a range of potential ap-
score would merit attention. In this re- plications of the approach, what we call
spect, possibility distributions have an Fuzzy Forecasting.
advantage over continuous probability APPENDIX:
distributions in that we can look back and FORMULA FOR THE SUCCESS SCORE
see what possibility values we gave to an Since we don’t know the exact height
outcome that actually occurred. Doing of the triangle at the breakeven sales of
so is not possible when working with 8,000, we proceed as follows:
continuous probability distributions, be- By definition, the height of the overall
cause each outcome has an infinitesimal triangle is 0 at the highest sales of 20,000
(virtually zero) probability of occurrence. and 1 at the most plausible sales of 8,000.
https://foresight.forecasters.org FORESIGHT 19
However, the shaded triangle has its To get its area, we multiply the height of
highest point at sales of 9,000. Moving the shaded triangle by half its base. Its
from right to left along the axis, sales of base is 20,000-9,000 = 11,000 and we
9,000 are (20,000-9,000)/(20,000-8,000) have (11/12) x 0.5 x 11,000 = 5,042 after
= 11/12 of the distance between 20,000 rounding.
with a height of 0 and 8,000 with a height Finally, the success score is the proportion
of 1. Therefore, the shaded triangle has a of the area of the triangle that is shaded,
height of 11/12 x 1 or 5,042/7,000 = 0.72
REFERENCES
Borges, R.E.P., Dias, M.A.G., Neto, A.D.D. & Meier, A.
(2018). Fuzzy Pay-off Method for Real Options: The
Paul Goodwin is an Emeritus Pro- Center of Gravity Approach with Application in Oilfield
fessor at the University of Bath, author of Abandonment, Fuzzy Sets and Systems, 353, 111-123.
numerous books and articles on the use Dubois, D. & Prade, H. (2011). Possibility Theory and
of judgment in forecasting, and Foresight’s Its Applications: Where Do We Stand, Mathware and
Editor for Hot New Research. Soft Computing, 18(1), 18-31.
Edwards, W. & Barron, F.H. (1994). SMARTS and
p.goodwin@bath.ac.uk SMARTER: Improved Simple Methods for Multi
Attribute Utility Measurement, Organizational
Behavior and Human Decision Processes, 60(3), 306-325.
Kuhn, T.S. (1962). The Structure of Scientific Revolutions,
Chicago: University of Chicago Press.
Steve Morlidge is founder of Catch
Norris, P. (2001). Apathetic Landslide: The 2001 British
Bull ( https://catchbull.com/ ), co-author
General Election, Parliamentary Affairs, 54, 565-589.
of Future Ready: How to Master Business
Forecasting (2010) and author of The Little Russo, J.E. & Schoemaker, P.J. (1992). Managing
(Illustrated) Book of Operational Forecast- Overconfidence, Sloan Management Review, 33(2),
ing (2018). He has contributed nearly 20 7-17.
articles for Foresight covering forecast ac- Shackle, G.L.S. (1961). Decision, Order and Time in
curacy, forecastability, inventory control, Human Affairs, 2nd edition, Cambridge University
and a multi-part article on “Guiding Principles of the Forecasting Press, UK.
Process” (2012). Shaheen, A.A., Fayek, A.R. & AbouRizk, S.M. (2007).
Fuzzy Numbers in Cost Range Estimating, Journal
steve.morlidge@catchbull.com
of Construction Engineering and Management, 133,
325-334.
Taleb, N.N. (2007). The Black Swan – The Impact of the
Highly Improbable, London: Allen Lane.
Zadeh, L.A. (1999). Fuzzy Sets as a Basis for a Theory
of Possibility, Fuzzy Sets and Systems, 100, 9-34.
20 FORESIGHT 2021: Q4