Two people research a stock. One checks original documents; the other uses chat messages and an AI summary. Later, one loses money and the other earns it. Which deserves the better process assessment? Withhold your answer: the most distracting information belongs at the end.
Read their decisions under shared constraints before seeing profit or loss. The purpose is to preserve useful practices and identify omissions that a gain cannot erase, rather than assign either person a permanent investor-quality label.
Define the assessment before guessing a winner
You will evaluate process and outcome separately, distinguish overconfidence, hindsight, and outcome bias, and keep a review that can recover original probabilities, horizons, and reasons for revisions. The cards teach a method; they are not psychological tests or estimates of real investment ability.
Both separate USD simulation accounts begin October 29, 2026 with $10,000 cash, no other holdings, borrowing, or external flows, and no use of essential household money. Both observe sixty trading days. Their securities are fictional, initially priced at $100, with a simulated full sale at the supplied sixtieth-day close.
Predefine two constraints: single-security notional no greater than $1,500, and loss under a 40% price-decline stress plus all fees no greater than $450. Each purchase and sale costs a fixed $1, with no tax, dividend, or other cost. Neither these budgets nor the 40% decline is a universal recommendation or maximum loss. Closing-price execution is a bookkeeping assumption, not a real-market promise.
The binary forecast event is a sixtieth-day close strictly above the starting $100, excluding fees. With complete prices, equal or lower means nonoccurrence. Fix price source, deadline, and event before observing results. Price-event success and a net account profit are not identical.
Card A: documented inputs do not validate a probability
A retains the synthetic company's complete revenue and cash-flow tables, publication dates, and locations for each number. Revenue grows 12% and operating cash flow is positive. Receivables grow 25%, and one major customer contributes 30% of revenue; both appear among the contrary considerations.
The proposed explanation is that a new product may improve revenue, with receivables and customer concentration posing failure conditions. In thirty supplied historical synthetic cases with the same horizon and price-event definition, eighteen events occurred. A uses 18/30 = 60% as a working probability, explicitly leaving sample selection, dependence, and transfer to different conditions unverified.
A also gives a subjective 80% prediction interval for the sixtieth-day price change: −25% to +25%. That is a working coverage claim, not guaranteed boundaries or another expression of a 60% chance of rising. Track interval coverage separately from binary-event occurrences.
The simulated position is ten shares, or $1,000 notional. The 40% stress loses $400 plus $2 round-trip fees, totaling $402, below $450. Notional is below $1,500. Entry costs $1,001, leaving $8,999 cash. Research evidence still does not justify real implementation, so actual new capital remains zero.
Every twenty trading days, A checks newly published evidence. A changed customer-concentration or cash-flow definition pauses additions and generates an update. The original sixty-day forecast still settles on schedule. An earlier trade change needs separate actual-action and original-forecast fields, rather than deletion of the forecast.
Card B: completing a form is not supporting its contents
B records the same date, event, and sixty-day horizon but forecasts a 90% chance of rising. Sources are forwarded chat and an AI summary without verified original links. Its claim of recent accuracy lacks failures, unresolved forecasts, and abandoned attempts.
B notices revenue growth without checking cash-flow tables, receivables, or customer concentration. Its subjective 80% interval is +5% to +25%, with no reference sample or construction rationale. A narrow interval, or one wholly above zero, is not automatically wrong. The problem is whether evidence supports the confidence expressed.
B proposes fifty shares, $5,000 notional. Entry costs $5,001, leaving $4,999. The shared 40% stress loses $2,000 plus $2 fees, or $2,002, above $450. Notional also exceeds $1,500. No exception is justified and the application is not withdrawn.
To examine a defective plan alongside its later outcome, the teaching scenario nevertheless executes those fifty simulated shares. This is not policy approval. A real process should reject the excessive request before expecting profits to supply missing authorization.
B's review note says a rise proves the model, while a decline means waiting until recovery to settle. That abandons the original horizon. These documented source and policy defects can be identified now without guessing whether B is confident, anxious, or deliberately misleading.
Save your first assessment before opening the results
Give each card zero or one on five checks: clear event and horizon; traceable original inputs; treatment of contrary evidence; compliance with amount and stress limits; and checkable updating and settlement arrangements. Attach a card location to every point and write one unresolved question. Save this version before disclosure.
Under the supplied conditions, A completes all five and B only the first: scores five and one. These are teaching-completeness checks, not a professional ability scale. A still needs scrutiny: were the thirty cases selected, do eighteen occurrences add independent information, and what supports the 80% interval?
A's organized format does not exempt the model from criticism. Its advantage is retaining uncertainties and limiting actual new funding to zero, not proving that 60% is accurate. B's probability field cannot replace source and constraint checks.
Open only after saving your first assessment
The scenarios deliberately reverse process and outcome. A's sixtieth-day close is $80. Selling ten shares produces $800 less $1; adding $8,999 gives $9,798. Net loss is $202, or −2.02%. The upward event did not occur, but the −20% price change lies inside A's −25% to +25% interval.
B closes at $120. Fifty shares produce $6,000 less $1; adding $4,999 gives $10,998. Net gain is $998, or 9.98%. The event occurred, and +20% lies inside B's interval.
A has a better documented process and an adverse outcome; B has known defects and a favorable outcome. Each is one observation, proving neither superior enduring ability nor permanent uselessness. Two intervals each covering once do not establish 80% calibration.
What can legitimately change after disclosure?
Add an outcome column recording simulated profit, trades, and event occurrence. Revisit the process scores. If you want to remove B's source and funding objections merely because it earned money, preserve that proposed reason for examination.
Outcome bias means outcome knowledge influences evaluation of the earlier decision's quality. Baron and Hershey's original experiments used uncertain medical and monetary decisions; favorable outcomes affected judgments of thinking quality.[1] Their tasks and participants are not every investor, and the experiments cannot diagnose everyone who reassesses a trade.
Do not reverse the error by declaring outcomes irrelevant to every process review. They can reveal unknown evidence about a data error, uncontrolled execution, or an omitted model factor. Identify the new fact that changes your understanding instead of exchanging one profit or loss for a quality label.
Suppose A had received a customer-cancellation notice before deciding but failed to read it. Discovery is new evidence of an earlier process defect, warranting a lower assessment and repaired checks. “Bad outcomes can be luck” does not excuse it. A previously listed adverse possibility occurring, by itself, needs more observations before settling model accuracy.
Hindsight reconstructs what was knowable
Hindsight bias concerns overstating an outcome's earlier foreseeability once it is known, remembering a tentative judgment as more certain. Outcome bias instead affects evaluation of how well the decision was made. They can coexist without being the same question.
Fischhoff's research compared probability and information-relevance judgments after different outcome reports, finding effects on reconstruction of earlier knowledge.[2] This does not make every claim of prior concern false. Compare it with dated records to determine what concern was actually expressed.
A's 60% cannot become nearly certain afterward. B's 90% cannot become just a casual thought when it fails. Updates may have new reasons, but the original remains recoverable. Where memory conflicts with documents, record uncertainty instead of asking AI to supply a convincing earlier thought.
Confidence must be compared with evidence
Here, overconfidence means confidence in accuracy or ability exceeds evidential support. High probability is not necessarily overconfidence; low probability is not automatic care. Examine the event, available information, and track record rather than diagnose tone, one position size, or one winner.
CFA Institute's public overview includes overconfidence and hindsight and treats recognition as an initial step.[3] The exercise uses observable checks rather than permanent personality labels or fixed psychological multipliers.
Success in one task also needs a transfer argument. Understanding a business does not establish short-term price forecasting. Identifying a trend does not establish fees, fills, or household suitability. Record competence within its actual tasks and inputs instead of replacing the limits with expert status.
Settle probabilities and ranges against their original questions
As in G13, calibration compares advance probabilities with frequencies of the corresponding events. Align definitions, horizons, information cutoffs, and settlement sources. Revenue growth and sixty-day price increases cannot share a denominator called successful predictions; unresolved and losing forecasts cannot vanish.
The Met Office's reliability diagrams group issued probabilities and compare observed frequencies.[4] Applying that record structure to investments is a teaching transfer, not weather evidence establishing financial skill. Dependence, changing conditions, and question selection remain issues.
Use a standalone synthetic card of ten comparable forecasts, all 60%. Occurrences in order are yes, no, yes, yes, no, no, yes, no, no, yes. Five of ten occur: 50%, ten percentage points below 60%. Ten records prove neither miscalibration nor stable ability.
Save the discrepancy and inspect sampling, rationale, and further comparable records. Small samples do not authorize deleting unfavorable observations. Base probabilities, information added by forecasts, and costs contribute more than hit counts alone.
Track 80% prediction intervals with coverage and width separately. An extremely wide interval may cover often but contribute little to decisions; a narrow one may understate uncertainty. One covered result for each card establishes only those two inclusions, not an 80% promise for later outcomes.
Apply a specific correction
Explanation question: What differs between outcome bias and hindsight bias?
Application question: A lost while respecting the simulation limits. Does that establish its 60% forecast was wrong? Write the missing reasoning before checking.
Outcome bias affects quality evaluation; hindsight affects reconstructed prior knowability. One nonoccurrence is compatible with the 40% alternative A already allowed. Inspect the sample and model without treating one failure as final refutation or five process points as immunity from criticism.
For a separate variation, an original table confirms A's receivables did not grow 25% year over year. A confused figures two years apart with a one-year comparison; the correct table and dates were available before deciding. Which record changes, and which versions must remain?
Add a documented sourcing/calculation defect and corrective action, rather than call it luck. Do not rewrite simulated trades or realized endpoints. Retain the original 60% and interval, then append corrected information, reasoning, and update time. Whether to restudy or continue zero real funding depends on the corrected full case, not an invented success rate inferred from one error.
Save a two-part review
The first part preserves then-available inputs, dates, event and horizon, probability/range and rationale, contrary evidence, amount limits, updates, and stopping conditions. The second adds outcomes, actual trades, fees, informational differences, newly evidenced process defects, and reasons assessments changed.
Retain scores before and after disclosure. Distinguish factual corrections from changes prompted only by profit or loss. Check that event and horizon stayed fixed, gaps were not erased, and original and revised forecasts remain traceable. Already-viewed outcomes or an unsaved first version mean retrospective reconstruction, not genuine blinded assessment.
Keep the process open to evidence-based criticism
Assess then-known inputs and constraints first, then add outcomes and new facts separately. Profits cannot wash away known defects, but losses cannot automatically be dismissed as luck. A justified revision identifies the evidence and its location.
Accumulate records under original definitions, horizons, probabilities, and ranges, preserving revisions and ability limits. One return cannot establish calibration or transform an incomplete procedure into skill. A review is useful when it identifies what should change and why.
References
- Baron and Hershey: Outcome Bias in Decision Evaluation. JPSP 54(4), 1988, 569–579. Original experiments; their participants and tasks are not identical to investing.
- Fischhoff: Hindsight ≠ Foresight. JEP: Human Perception and Performance 1(3), 1975, 288–299. Outcome knowledge and reconstruction of earlier judgments.
- CFA Institute: The Behavioral Biases of Individuals. Public 2026 overview and summary, not the paid curriculum.
- Met Office: Reliability and Sharpness Diagrams. Probability-group and observed-frequency interpretation; financial application is this book's teaching transfer.
Sources checked October 30, 2026. Cards, trades, price paths, and calibration sequence are synthetic, not an actual psychological or financial experiment. AI-assisted draft; independent AI full-text and bilingual review is complete. Human financial review and target-reader testing have not been conducted and are not scheduled for this production round.