Negative Prediction Error Review¶
Review — instantiates Prediction-Error Learning Calibration
A recurring review of outcomes that came in worse than expected — separating genuine negative surprises worth learning from, from failures that were already priced in, without turning the lesson into punishment.
Not every bad outcome is a lesson, and not every lesson should be a punishment. Negative Prediction Error Review is the recurring review that isolates the worse-than-expected outcomes — the negative tail of the prediction error — and asks of each whether the badness was genuinely surprising or merely a risk that was already known and priced in. Its defining discipline is twofold: it deliberately handles only the negative-signed errors (leaving the upside to its twin), and it polices the safety of learning from them — because negative signals, applied carelessly, teach fear, concealment, and scapegoating rather than improvement. A loss that fell within its expected range carries no new information and gets set aside; a loss that broke expectation gets studied. The review exists so that an organization learns from real downside surprises without over-punishing the expected failures — which is how you teach people to hide risk instead of report it.
Example¶
A fund runs a quarterly loss review — but pointedly not a review of every position that lost money. It looks only at positions that lost more than the model expected them to. A holding that fell squarely within its forecast downside band is passed over: expected, no lesson. One that broke its downside expectation gets dissected for what the model missed. And the review is explicitly blameless about the decision versus the outcome: the analyst who sized a position correctly on the evidence available, and was then blindsided by a genuine surprise, is not punished for the loss — because punishing sound decisions for unlucky results just teaches everyone to stop taking well-judged risks and to bury the ones they do take.[n1] What comes out is a short list of real model failures and zero fear-driven incentives — which is exactly the split the review exists to produce.
How it works¶
- Filter to the negative tail. Take only outcomes worse than expected by more than a set threshold; ordinary and expected-bad results are excluded from the start.
- Separate surprising-bad from expected-bad. Judge each against its prior expectation, not its raw badness, so priced-in losses are forgiven and genuine misses are surfaced.
- Extract the lesson from the genuine surprises. Only the true negative surprises feed a model update.
- Safety-check the incentive. Examine what the punitive signal teaches — whether it drives improvement or drives fear, concealment, and blame — and adjust before it hardens.
The distinguishing feature is the pairing of sign-filtered review with an explicit safety check on punishment.
Tuning parameters¶
- Surprise threshold — how far below expectation an outcome must fall to earn review. Tight catches subtle misses but floods the agenda; loose focuses attention but lets moderate surprises slip.
- Expected-bad tolerance — how much priced-in badness to forgive outright. Generous protects honest risk-taking; stingy risks punishing sound bets for their outcomes.
- Punishment stance — where the review sits between pure learning and accountability. This is the ethical dial: lean too far toward accountability and honesty collapses; too far toward learning and genuine negligence is excused.
- Cadence — how often the review runs, trading freshness of memory against reactive churn.
- Scope — whether it hunts individual decisions or systemic causes.
When it helps, and when it misleads¶
Its strength is that it stops two distinct errors: over-punishing expected failures (which teaches people to conceal risk) and overreacting to a single unlucky draw as if it disproved a sound approach. By grading the decision against what was known rather than against the loss, and by keeping the process blameless, it extracts the real downside lessons while protecting the honesty the system runs on.
Its failure modes are the ones every post-mortem faces. Hindsight makes every loss look predictable, so the review must actively resist rewriting expected-bad as should-have-known. Left unchecked it curdles into blame theater — a search for someone to punish rather than something to learn. Its classic misuse is convening it to justify a verdict already reached about a person or a bet. The discipline that guards against this is the blameless post-mortem stance: judge the decision on the information available when it was made, keep the focus on the model and the system, and treat the punitive dimension as a hazard to be managed, not a goal.
How it implements the components¶
Negative Prediction Error Review realizes the downside half of the archetype's error handling, plus the safety of learning from punishment:
positive_negative_error_separation— it operationalizes pulling the negative-signed errors out for their own handling; the positive tail is routed elsewhere rather than blended in.ethical_reward_safety_review— it checks that the punitive signal drives improvement rather than fear, concealment, or scapegoating, and adjusts the incentive before it hardens.
It does not harvest the upside surprises (that is Positive Surprise Capture, which handles the opposite sign of the same split), record the raw deltas it filters (Prediction–Outcome Delta Log), or set how hard to update on a confirmed negative surprise (Learning Rate Schedule).
Related¶
- Instantiates: Prediction-Error Learning Calibration — the review is the archetype's disciplined handling of worse-than-expected outcomes.
- Consumes: Prediction–Outcome Delta Log supplies the negative-signed entries the review draws from.
- Sibling mechanisms: Positive Surprise Capture · Prediction–Outcome Delta Log · Learning Rate Schedule · Credit Assignment Trace · Reward Signal Red Team · Expectancy-Calibrated Feedback Form · Calibration Curve Review · Reward Baseline Dashboard · Shortcut Probe Holdout Set · Surprise Threshold Alert · Temporal-Difference Update Rule
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Negative Prediction Error Review operates as a bounded evaluation of existing evidence or work that produces a finding or disposition because it a recurring review of outcomes that came in worse than expected — separating genuine negative surprises worth learning from, from failures that were already priced in, without turning the lesson into punishment.
Independent corroboration: The frozen evidence defines Negative Prediction Error Review as 'A recurring review of outcomes that came in worse than expected — separating genuine negative surprises worth learning from, from failures that were already priced in, without turning the lesson into punishment', so its operative form is Assessment, Review & Assurance.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Psychology
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Psychological work on prediction error, outcome bias, and counterfactual learning supplies the review's distinction between surprise and a merely bad result.
Related originating lineages:
- Behavioral Economics — Decision research developed calibration against prior expectations and analysis of choices under uncertainty.
- Cognitive Science — Belief-updating models materially shape the mechanism's calibration of lessons to prior expectations.
- Organizational & Management Science — Management practice contributed the recurring, nonpunitive review that converts forecast misses into organizational learning.
Review resolution: Authoritative-source research resolves the primary-origin disagreement. Prediction error as the mismatch between expected and actual outcome is a psychological learning construct; the recurring nonpunitive review is an organizational synthesis over it. Origin breadth is limited to formative lineages; present-day applicability is recorded separately as domain_reach=multi_domain.
Attribution caveat: Prediction error is a psychological construct, but the named mechanism is an organizational review ritual.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; medium confidence.
Sources consulted:
Notes¶
This review is one half of a matched pair with Positive Surprise Capture: the same positive/negative split, opposite sign. Running only the negative half is how an organization becomes loss-obsessed — endlessly dissecting what went wrong while never studying what went unexpectedly right — so the pair is meant to be run together.
[n1] Outcome bias — judging a decision by how it turned out rather than by what was known when it was made; Annie Duke calls the error resulting. Grading a decision against its prior expectation, not against the loss, is the guard this review is built to enforce. ↩