From 1% to ‘Above 50%’: How Much Action Can Probability Demand Before the Future Happens?

Geoffrey Chen

When Numbers Start Making Decisions · Season Five, “Things That Have Not Happened Yet” · Article 12

1. Season Five has not studied one kind of probability

A 30% chance of rain refers to measurable precipitation at a place and time. A 1% AEP flood is an annual chance that a magnitude will be reached or exceeded. BAL-29 and BAL-40 are not occurrence probabilities; they are attack categories in a design bushfire scenario. Category 2 and 3 tropical cyclones are principally classified by wind.

The electricity system’s 0.002% is annual expected unserved energy. Four weeks of income is a behavioural threshold that turns a modelled premium into an uninsured scenario. FIT’s 20 micrograms is a faecal-haemoglobin triage line. NIPT’s roughly 99% is a detection rate for trisomy 21, not the positive predictive value for one pregnancy.

DNA’s 99.9% exclusion supports a source comparison, not a probability of guilt. A p-value below 0.05 describes compatibility with a specified statistical model, not truth. The national terrorism level’s “above 50%” is a national twelve-month intelligence assessment of an attack or attack planning.

These numbers have different mathematical meanings. They share an institutional function: before an event has occurred, a diagnosis exists or a conclusion is established, they give present action a structure.

The central question of Season Five is how much preparation, testing, investment, restriction or coercion uncertainty can legitimately trigger.

2. First, write the complete event

Any probabilistic decision should complete one sentence:

At what place, over what time, under what conditions and model, what event has what probability?

“Thirty per cent rain” is incomplete without 0.2 millimetres, the period and location. “One in 100 years” is incomplete without annual exceedance and model version. “Terrorism above 50%” is incomplete without the national scope, twelve months and the combined attack-or-planning event.

Many conflicts arise because people are discussing different events. Insurers discuss annual property loss while residents ask whether a community remains liveable. An NIPT laboratory describes detection while a pregnant person asks about one pregnancy. A DNA scientist discusses trace source while a jury judges an offence.

An incomplete event should not acquire major decision authority.

3. Separate the prediction from the action threshold

A prediction asks what might occur. An action threshold asks what we will do under a given state of evidence. Institutions frequently conceal both inside one number.

The Bureau publishes a rain probability but does not cancel an event. It classifies a cyclone while emergency agencies decide evacuation. AEMO forecasts USE while rules and governments determine reserves. NIPT high risk opens diagnostic choice without completing an irreversible pregnancy decision.

Every action boundary needs a responsible actor. Whoever connects a number to stopping work, refusal, investigation or force must explain error costs, alternatives and legal authority. A normative decision cannot hide behind the phrase “the model recommended it”.

The same probability can therefore support different actions. A 30% chance of light rain may justify protecting equipment and not cancelling a city event. A 1% annual flood risk may justify higher floors in new housing but not immediate relocation of every existing community.

4. Consider probability and consequence together

Low probability does not mean no action, and high probability does not automatically legitimise force. At minimum, decisions need both likelihood and severity.

High-probability, low-consequence events usually invite routine preparation. Low-probability, high-consequence events justify precaution, redundancy and monitoring. High likelihood and high consequence may require rapid measures. Low likelihood and low consequence may be observed.

The matrix remains incomplete without reversibility, distribution and action cost. Irreversible environmental harm, loss of liberty and pregnancy decisions require thicker evidence than installing a marquee or arranging backup power, even at the same probability.

A general proportionality principle follows:

The more uncertain the number, the more reversible the action should be. The more severe the possible consequence, the less an institution can wait for certainty. The more coercive and irreversible the action, the stronger its reasons, review and exit conditions must become.

This prevents precaution from expanding without limit and prevents uncertainty from becoming a permanent excuse for delay.

5. Distinguish opening an opportunity from closing a right

Screening and probability are well suited to initiating further steps. Positive FIT can open colonoscopy. High-risk NIPT can open diagnosis. A DNA match can open an investigation. A flood map can open engineering assessment.

Using the same number to close rights requires more. A negative FIT cannot overrule symptoms. NIPT alone cannot support an irreversible pregnancy decision. Source-level DNA cannot complete conviction. A national threat level cannot reduce legal proof against an individual.

This is an opening–closing asymmetry. A broad, sensitive tool can reasonably open more doors when later stages confirm the result. It should not use the same coarseness to shut someone out permanently.

Policy follows the same logic. Four weeks of income can identify regions needing help; direct eligibility needs a further justification for people around the line. BAL can intensify approval but cannot guarantee or deny the fate of one house.

6. Divide error by direction and level

“The model may be wrong” is too vague. At least six questions are separate:

  1. Are the input data accurate?
  2. Does the number measure the object the decision actually concerns?
  3. Is the model calibrated in comparable cases?
  4. Is the action threshold placed reasonably?
  5. Are consequences proportionate to evidence?
  6. Can an affected person challenge the result and recover rights?

False alarms and misses must also be distinguished. A weather false alarm may mean carrying an umbrella; a missed cyclone or security threat may cost lives. Increasing medical-screening sensitivity increases false positives and diagnostic demand; reducing it misses disease. No overall accuracy score can allocate these errors for us.

Group averages hide where errors fall. Regional USE can be tiny while a long shortage harms vulnerable households. National insurance stress concentrates in remote and low-income places. Security measures may burden particular communities disproportionately.

7. Put time inside the decision

Future-facing numbers update. Weather forecasts change hourly. Cyclones intensify. Flood maps and BAL assessments change with evidence and landscape. Electricity projects are delayed. Diagnosis supersedes screening. A threat level has a twelve-month horizon.

Every predictive measure should include a next review, escalation condition, de-escalation condition and exit. Precaution without exit becomes permanent power. Observation without escalation becomes delay.

Action should be staged. Take low-cost, reversible steps first and preserve future choice. Escalate when evidence strengthens or time shortens. In infrastructure, that means scenarios, triggers and staged investment. In individual diagnosis, screening, review, confirmation and decision cannot be collapsed.

Evaluation must use the information available at the time. A risk that fails to occur because action prevented it is not automatically a false forecast. Neither can every unrealised prediction be excused as successful prevention. Version, counterfactual and action records make learning possible.

8. Explain uncertainty rather than adding a disclaimer

“Model for guidance only” explains nothing. Minimum communication names the object, denominator, period, model version, main assumptions, omissions, intended use and the person or body empowered to override it.

Across fields, the same compression errors recur: 30% becomes area or time; 1% AEP becomes a 100-year countdown; BAL becomes survival probability; 99% detection becomes individual diagnosis; 99.9% exclusion becomes guilt; p<0.05 becomes truth; a national 50% assessment becomes personal risk.

These are not merely failures of public intelligence. Institutions routinely hide full conditions behind a compact interface. The duty to explain rises with consequence. Experts need not expose classified intelligence or proprietary code, but affected people need to know what the number supports, what it does not and who reviews it.

9. Prediction participates in forming the future

Season Five’s deepest finding is that probability is not a mirror outside reality. Flood maps change buildings. BAL changes vegetation and siting. Reliability forecasts attract investment. Insurance stress testing prompts adaptation. Threat levels change protection and adversary behaviour.

Structures persist through prediction, action, feedback and correction. Future risk gains practical meaning in relations among data, models, institutions and response.

This does not justify returning to private intuition. Nobody possesses the atmosphere, grid, genome, national threat and every household’s risk. Knowing must be delegated to models, laboratories and institutions.

Delegating knowledge does not delegate responsibility. A system can calculate probability but cannot carry the reasons for cancellation, public spending, invasive testing, loss of liberty or emergency powers. After calculation, a person or institution still has to sign, explain, review and repair.

10. Ten questions for acting under uncertainty

For any future-facing number, ask:

  1. Event: What complete event does the number describe?
  2. Scope: What are the place, period, denominator and jurisdiction?
  3. Source: Which data, model, expert judgement and version produced it?
  4. Omission: Which hazards, people or behaviours remain outside it?
  5. Actor: Who connects prediction to a real decision?
  6. Proportionality: Do probability, consequence, cost and action intensity fit?
  7. Reversibility: Can an adjustable step preserve choice?
  8. Error: Who carries false alarms and misses?
  9. Update: What triggers review, escalation, de-escalation and exit?
  10. Remedy: Can rights and relationships be restored and the system changed?

The questions do not turn every field into one formula. They stop numbers crossing their evidentiary boundaries unnoticed.

Conclusion: we need a corrigible present, not a certain future

Probabilities and risk levels can legitimately trigger preparation, monitoring, investigation, diagnosis and staged investment. For severe low-probability harm, uncertainty is not a reason to wait until action is useless.

A single prediction should not independently cause irreversible personal deprivation or final judgement. The more coercive the action, the greater the need for specific evidence, independent review, time limits and remedy. A national or population probability cannot lower an individual standard of proof.

Season Five can be condensed into three sentences:

The more uncertain the prediction, the more reversible the action should be. The more severe the possible consequence, the less an institution can wait for complete certainty. The greater the coercion, the clearer the reasons, review and exit must become.

Modern institutions cannot know the future before acting. They can ensure that present action keeps the capacity to learn: to change direction when evidence arrives, restore relationships when forecasts fail and retain accountable commitment when a number gains authority.

Primary sources

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.