A prediction is a claim about a future, unseen or not-yet-observed state.
Weather forecasts, exam-score estimates, demand forecasts, failure predictions, economic projections, model classifications and expected delivery times all make statements about what should be observed later or elsewhere. They differ in target, horizon, method, probability, conditionality, calibration and consequence.
Quick answer: how should predictions be categorised?
- Target: what future or unseen state is being predicted?
- Horizon: seconds, days, years?
- Method: rule, model, expert judgement, trend, simulation?
- Form: point, range, category, ranking, probability?
- Conditionality: unconditional or dependent on assumptions?
- Probability: deterministic or probabilistic?
- Uncertainty: what range of outcomes remains plausible?
- Calibration: do predicted probabilities match observed frequencies?
- Validation: how has predictive performance been tested?
- Outcome: what was eventually observed?
This article complements How to Categorise Models and How to Categorise Scenarios: models may generate predictions, while scenarios explore possible worlds without necessarily assigning one forecast as expected.
1. Separate prediction from scenario
A prediction states what is expected or estimated to occur. A scenario describes what could occur under a set of assumptions.
2. Separate prediction from projection
A projection often extends a trend or model under explicit assumptions without claiming that those assumptions are the most likely future.
3. Separate prediction from explanation
A system can predict accurately without explaining mechanism, and explain well without predicting every instance precisely.
4. Target identity should be explicit
Predicting category, quantity, event, rank, duration or state requires different evaluation.
5. Point predictions give one value
They are compact but can hide uncertainty.
6. Range predictions preserve uncertainty
Intervals communicate a set of plausible values rather than one falsely exact estimate.
7. Categorical predictions choose a class
Pass/fail, high/medium/low and fault type are examples.
8. Probabilistic predictions assign likelihood
They preserve uncertainty directly instead of forcing one class too early.
9. Ranking predictions order alternatives
Search, recommendations and prioritisation may care more about order than exact probabilities.
10. Horizon changes difficulty
Longer horizons usually expose predictions to more unknown events and behavioural adaptation.
11. Short-horizon predictions can depend on current state
Nowcasting and immediate forecasting often rely heavily on fresh signals.
12. Long-horizon predictions depend more on assumptions
Demography, policy, technology and behaviour can dominate outcomes over years or decades.
13. Conditional predictions preserve assumptions
“If demand stays constant, stock lasts six weeks” should not be simplified to “stock lasts six weeks”.
14. Unconditional predictions imply stronger commitment
They should be used carefully when hidden assumptions remain material.
15. Rule-based predictions use explicit logic
Thresholds and decision trees can make outcomes easy to explain.
16. Statistical predictions estimate from patterns
They infer relationships from data and quantify uncertainty to varying degrees.
17. Mechanistic predictions come from system models
They derive expected outcomes from represented causal or physical relationships.
18. Expert predictions use judgement
They can integrate tacit knowledge but should preserve the expert, rationale and confidence.
19. Trend extrapolation assumes continuity
It can fail badly when structural breaks occur.
20. Ensemble predictions combine models
Combining independent models can reduce reliance on one model’s specific errors.
21. Prediction confidence is not prediction quality
A confident prediction can be wrong; quality must be evaluated against outcomes.
22. Calibration evaluates probability honesty
Events predicted at 70% should occur about 70% of the time across a suitable set of forecasts.
23. Accuracy and calibration differ
A system can classify many cases correctly while still assigning badly calibrated probabilities.
24. Discrimination measures separation
Some prediction tasks require distinguishing high-risk from low-risk cases more than estimating exact probability.
25. Error costs should shape evaluation
False alarms and missed events can have very different consequences.
26. Rare-event prediction needs special care
High overall accuracy can be meaningless when the important event is uncommon.
27. Prediction intervals need coverage testing
An interval labelled 90% should capture the eventual outcome at roughly the promised rate over suitable cases.
28. Structural breaks invalidate old patterns
Policy change, technology, crises and behavioural shifts can make historical relationships unreliable.
29. Prediction drift should be monitored
Performance can decay even when model code has not changed.
30. Outcome comparison closes the loop
Store what was predicted and what later occurred so performance can be audited.
31. Predictions need provenance
Record model, data, rule, expert, assumptions and version.
32. Revised predictions should preserve history
Do not overwrite yesterday’s forecast with today’s update if later evaluation depends on what was actually known then.
33. AI-generated predictions need task-specific testing
Fluent output is not evidence of forecasting skill; performance must be measured on held-out outcomes.
34. A practical prediction record
- prediction ID;
- target;
- prediction form;
- value, class or distribution;
- horizon;
- method;
- assumptions;
- confidence or probability;
- uncertainty interval;
- model or expert source;
- version;
- issued time;
- observed outcome;
- error;
- calibration status.
35. Prediction categories should improve evaluation
A useful scheme tells us which scoring method, uncertainty representation and review process fits the forecast.
36. The deeper idea
A prediction is a claim sent forward in time. Its quality becomes knowable only when reality returns an outcome.
To categorise a prediction well is to preserve what was predicted, how far ahead, by what method, with what uncertainty, and how the eventual outcome compared with the forecast.
Final answer
Categorise predictions by target, horizon, method, form, conditionality, probability, uncertainty, calibration, validation and outcome. Keep predictions separate from scenarios and explanations, preserve assumptions and versions, and always compare forecasts with what later occurred.