This requires a perspective on measurement that makes the concept of “uncertainty” forefront.
We often use instruments to measure things. For instance, a ruler to measure the size of a table. However, we rely on estimation when instruments do not exist for our area of measurement.
As an example, any weight scale you might own today is likely calibrated to an approximation of many intermediary approximations to an international prototype stored in an underground vault in France. The definition of a “kilogram” has changed over time. Only recently (Nov 2018) has it been defined by a universal constant. Even so, devices will be calibrated to an approximation of this constant.
In risk, we are concerned with future events and their impacts. No instrument exists that can directly measure future events. As a result, we often find ourselves needing to approximate the likelihood and impact of any number of potential future outcomes of an event by relying on expert interpretation of historical data, reference classes, and statistical models.
This approximation of future outcomes is typically called a forecast. As individuals concerned with future, undesirable events (our risk), we find ourselves forecasting the likelihood or impacts of these events.
As the role of information becomes more prevalent in a forecast, we can reduce (but never eliminate) our uncertainty in a given scenario.
The primary feature of this documentation is to make all concepts of “risk” subject to quantitative techniques, and forecasting is one of these important methods.
Some thoughts on definitions:¶
By sticking to principles (See: Principles), this documentation is opinionated on the usage of some risk language. Estimations are a form of approximation for any unknown value. A forecast is a an estimation of a value that doesn’t exist yet. An estimate is not necessarily a forecast, but a forecast is an estimate.
For the purposes of this documentation, there is not much different between something that is unknown (or, “yet to be revealed”) and something that hasn’t happened (“a future event”). For instance, an unknown quantity may also be a future value. It may not be known if a value exists yet.
An example: “Monsters are underneath the bed” could both be a future event (they might not be there yet, but they could be there soon) and also information to be revealed (they were / are always there, you just haven’t looked yet).
There are likely more grey areas for these terms. These will be worked out as opportunities to simplify arise.
Lastly, a prediction does not necessarily mean a “100% belief”, but that language should probably be avoided as it can be interpreted poorly. Forecast seems to be more appropriate, given people’s familiarity with the uncertainty of weather predictions. IE, “We predict an 80% likelihood” versus “We forecast an 80% likelihood”.
Forecasting is a disciplined practice to estimate the likelihoods and impacts of future events. It is a subject matter with over a half-century of multidisciplinary research into risk, decision making, and predictions.
Every day, you forecast things related to your basic needs. Choosing what time you wake up in the morning is a forecast of how much time you’ll need to prepare for the day. Peeking out your window informs a forecast of the weather which informs a decision about clothes to wear. Examples are infinite.
You’ll notice that you can seek information to support these forecasts, outside of relying on your expertise or intuition. Sometimes this data is readily available, sometimes it is not.
Forecasting is relied on in either case. Even when a meteorologist is presented with significant data about a ten day forecast, it is common for them to make a personal estimate based on that data instead of using it directly.
With forecasting, we do our best to approximate values that best represent our intuition. We stick to methods that improve our forecasting skills, and defend ourselves against well known forms of cognitive bias. As subject matter experts of the risks care about, we can create risk data that is highly leveraged by quantitative methods.
When culturally supported and invested in, groups of engineers can attack large risks methodically.
Forecasts often include a value of “confidence” associated with them. For instance, “I am 50% sure it will rain tomorrow.” would indicate that the forecaster will be historically wrong in half of the instances where they’ve made a 50% claim.
If someone is 99% certain, their track record would be incorrect one in one hundred cases.
This makes the values of 0% and 100% very special, as they would indicate that the forecaster expects a perfect track record.
You do not know anyone with a perfect prediction track record.
Volumes of research show that humans are poorly calibrated without training and practice. An uncalibrated individual may frequently use the phrase “I’m 90% sure” and display a track record of being far worse, as an example.
Research shows that individuals can be very easily calibrated with minimal training, and regular practice supports this as well. (See: Tetlock)
Forecasts that include their associated confidence can make use of the Brier Score to record accuracy over time. This is simply calculated as the “Squared Error”.
The Good Judgement Open has an accessible definition of the Brier Score:
The Brier score is the squared error of a probabilistic forecast. To calculate it, we divide your forecast by 100 so that your probabilities range between 0 (0%) and 1 (100%). Then, we code reality as either 0 (if the event did not happen) or 1 (if the event did happen). For each answer option, we take the difference between your forecast and the correct answer, square the differences, and add them all together. For a yes/no question where you forecasted 70% and the event happened, your score would be (1 – 0.7)2 + (0 – 0.3)2 = 0.18. For a question with three possible outcomes (A, B, C) where you forecasted A = 60%, B = 10%, C = 30% and A occurred, your score would be (1 – 0.6)2 + (0 – 0.1)2 + (0 – 0.3)2 = 0.26. The best (lowest) possible Brier score is 0, and the worst (highest) possible Brier score is 2.
An average Brier score is useful for tracking the reliability of a forecaster. It can be tracked by certain topics, panels, individuals, etc.
For instance, let’s take a batch of some pretty good weather predictions.
|Forecast||% Rain||% No Rain||Outcome||Brier Score||Brier Score (Work)|
|3||0.334||0.666||No Rain (0)||0.223112||
|4||0.01||0.99||No Rain (0)||0.0002||
This table shows an average Brier Score of
0.0617024. If we observed this forecast score from our local meteorologist, we’d be pleased and consider this forecast source useful. Let’s put together a table of pretty terrible weather forecasts for comparison.
|Forecast||% Rain||% No Rain||Outcome||Brier Score||Brier Score (Work)|
|3||0.77||0.23||No Rain (0)||1.1858||
|4||0.88||0.12||No Rain (0)||1.5488||
This table shows an average brier score of
1.49556. Any reasonable individual would consider those forecasts not useful.
Your industry will vary on what a “useful” threshold for a forecast source would be. For instance, a Brier Score that forecasts data related to part failures and explosions will be very different from a risk forecast about missed project deadlines. This documentation leaves that up to the engineers involved to set their requirements.
However, all industries can agree that engineers seeing a reduction of a Brier Score over time is a favorable trend, and is a useful engineering metric that can be targeted over time and improved upon.
Forecast sources can also be compared with the “Brier Skill Score”, in which we can discover better risk prediction models or methods. This is heavily used in meteorology to compare the value of a predictive model to a tried and true model, like a simple historical average. It is expressed simply with two Brier scores being compared below.
BrierSkillScore = 1.0 – BrierScoreNew / BrierScoreReference
Types of Outcomes¶
A scenario can prompt for several types of outcomes to forecast. Depending on the risk you are hoping to measure, you may want to prompt an expert for a different type of outcome.
Yes or No, Over / Under, and Multiple Options are probability distributions. They can be used to forecast with a percentage likelihood that a certain event will, or will not happen. Likelihood is split between mutually exclusive options, and must equal 100%.
Credible intervals are a bit different. They can be used to forecast an unknown value, like the potential impact (money lost, injuries, delays) associated with any scenario.
Yes or No¶
The simplest type of forecast asks an expert for their belief of a binary outcome. For instance:
Will it rain tomorrow?
(Yes / No)
A forecaster may express themselves by saying Yes: 60%, No: 40%, if they believe it’s more likely that not to rain. Or for instance, Yes: 0.01%, No: 99.99% if the forecaster lives in the desert.
Both likelihoods would need to sum to 100%.
Over / Under¶
To include some aspect of “impact” in a risk, you can bake an over / under value into the scenario.
Will there be more than **three inches** of rainfall tomorrow? (Yes / No)
(Yes / No)
Both likelihoods would need to sum to 100%.
This is similar to the previous forecast, but instead adds a numeric condition that must be met. This is useful when investigating the likelihood that some risk will meet a threshold or tolerance level you need to better understand. For instance, there may be a legal reason to close down schools with a certain height of snow, or maybe a certain amount of losses that your insurance couldn’t cover.
Alternatively, this could help determine a value for parametric insurance, in which a payout occurs if a threshold is met. For instance:
A policy that pays $100,000 if an earthquake with magnitude 5.0 or greater occurs.
Some forecasts may include many outcomes. For instance:
Our potential customer has decided on a vendor.
This could be answered with multiple options, like (A: Us, B: Competitor 1, C: Competitor 2, D, Competitor 3, E: No Decision / Walkout.)
% Likelihood A: Us B: Competitor 1 C: Competitor 2 D: Competitor 3 E: No Decision / Walkout / Other
All likelihoods would need to equal 100%.
A credible interval represents a range of possible values, and also includes a percentage belief (
confidence) that the outcome will fall into it. A forecast source (a model, or an expert) would then expand their range of values to increase their expression of uncertainty, and increased effort and data would widen or narrow this uncertainty. For example:
Police have responded to a protest at City Hall.
(# of arrests, 70% confidence)
A forecast source may answer this with an interval of 5-10 arrests, with a caveat that they expect with 70% likelihood, to eventually be correct (their confidence). If, for instance, they were asked for a less uncertain forecast, they may respond with a 6-8 interval with a 50% confidence.
Depending on your subject matter, it should be clear that some combinations of confidence and uncertainty are more or less useful than others. For instance, a 50% confidence of -1000-1000 arrests is not very useful, given the scenario of arrests at city hall.
A visual example of a percentage belief that an unknown value will end up within this range when revealed.:
70% Certainty │ │ │ │ │ ▼ 5 10 ▽──────────────▽ ◀─────────────────────────────────────────────────────────────────────▶ ... -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ...
To summarize, a forecaster would provide:
- An interval (min-max)
- A percentage belief the outcome lies within.
A scenario can also demand the percentage belief beforehand.
Divide and Choose¶
Divide and choose is a mental heuristic to determine if odds are fair or not. It is similar to the children’s “fairness” concept where one child slices a piece of cake, and another child chooses the slice they’d like.
This method prevents the first child from slicing unevenly and taking the larger piece.
This equates to forecasting, where instead of assigning “fair odds” for an event, a forecaster assigns an extreme likelihood to a scenario in pursuit of a stronger accuracy score.
As forecasting can often be related to gambling or a decision market, it can appear advantageous to “win” a forecast and aggressively assign likelihood to one option or another. A goal of forecasting is to assign “fair odds” that represent the whole uncertainty associated with an event or value, instead of strong accuracy scores.
Strategies and incentives to maximize accuracy scores over calibration can hinder this approach, as it is not meant to be “gamified”.
Principle of Indifference¶
The principle of indifference is a rule of thumb that divides a likelihood across all of its options. For instance, 50/50% or 25/25/25/25%.
When faced with these odds, a forecaster may find themselves disagreeing with them. If this is the case, it’s likely that the forecaster has opinions they may express numerically.
The Absurdity Test¶
The absurdity test assigns extreme and irrationally formed likelihoods or values to a forecast, testing the opinions of a forecaster. For instance, “A small child can eat between zero and one million pies in a sitting.”
When faced with such a test, a forecaster may be encouraged to start making a forecast less absurd. For instance “Well, a child can at least eat half of a pie, and maybe up to five pies, in extraordinary circumstances.”
This form of test has been used as an interview prompt in psychological research since the 1900’s.
When data is not available to study a risk, alternative data may suffice as a reference. For instance, the history of reversals in the Supreme Court may inform a type of case that may be considered unprecedented.