Cherry-picking a data cluster and building a story around it after the fact, ignoring everything that doesn't fit.
Imagine you throw a whole bag of darts at a wall with your eyes closed. Then you walk up, find the three darts that happen to be close together, draw a circle around them, and tell your friends you're an amazing dart player. You're not lying that those darts are close — but you're hiding all the other darts that went everywhere. That's what this fallacy is: picking the part that looks good and pretending the messy rest doesn't exist.
The Texas Sharpshooter Fallacy occurs when someone sifts through a large or complex dataset, identifies an apparent pattern or cluster after the fact, and then presents it as though it were a predicted or meaningful finding — while discarding all the surrounding data that fails to support the conclusion. Unlike forming a genuine hypothesis before examining evidence, this fallacy reverses the scientific process: the 'target' is drawn around whatever already looks interesting. It is especially pernicious because the patterns discovered are often real in a superficial statistical sense — they just aren't meaningful, having emerged from random variation in sufficiently large data. The fallacy is a core driver of the replication crisis in science, where researchers engage in 'HARKing' (Hypothesizing After Results are Known) and selectively report only the variables that yielded statistical significance.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors and analysts frequently back-test trading strategies across decades of market data, then highlight the specific combination of indicators that would have yielded the highest returns — without accounting for the thousands of indicator combinations that failed. This leads to overfitted models that perform well historically but collapse on new data.
Epidemiologists face constant pressure from communities reporting apparent disease clusters. When boundaries are drawn around observed cases after the fact rather than defined in advance, nearly any region can appear to have an abnormally high rate of illness, leading to false alarms, wasted resources, and public fear about non-existent environmental hazards.
Administrators may examine student performance data across many variables — teacher, classroom, time of day, curriculum version — and highlight whichever combination shows the best results as a 'proven approach,' ignoring the many other combinations that showed no effect and failing to validate the finding with new data.
People sometimes review a partner's past behavior after discovering infidelity, selectively stringing together small moments that 'now make sense' as evidence of a pattern of deception, while ignoring the overwhelming number of interactions that were genuinely honest and loving.
Product teams running multivariate A/B tests across many user segments, features, and time windows will inevitably find some combination that shows a statistically significant result. Treating that finding as a validated insight without replication leads to feature changes that don't actually improve the product.
HR departments may identify that several top performers share a particular trait — such as attending the same university or having a hobby in common — and build hiring criteria around it, ignoring the many employees with the same trait who performed poorly and the top performers who lack it entirely.
Pundits cherry-pick economic indicators, crime statistics, or polling data from specific time windows or regions that support their preferred political narrative, while ignoring the broader dataset that tells a more ambiguous or contradictory story.
Epidemiologist Seymour Grufferman, 1977. First appeared in a paper on Hodgkin's disease clustering (Cancer, 39: 1829–1833), using the sharpshooter metaphor to warn against post-hoc interpretation of disease clusters. The concept gained broader traction in critical thinking and statistics during the 1990s.
Ancestral humans who over-detected patterns — seeing a predator in rustling grass even when it was just wind — survived more often than those who under-detected. This asymmetry in the cost of errors (Type I errors being cheaper than Type II errors) selected for brains that aggressively find clusters and assign meaning, even at the expense of frequent false positives.
Machine learning models are highly susceptible to this fallacy during training. When models are evaluated on the same data used to find patterns (overfitting), they 'draw the bullseye' around noise in the training set. Feature selection algorithms that test thousands of variables and report only the significant ones without correction replicate the fallacy at computational scale. In LLMs, cherry-picked benchmark results or selectively reported evaluation metrics can create a misleading picture of model capability.
50 cards are free to preview. Buyers unlock the rest of the deck plus the interactive training — Spot-the-Bias Quiz unlimited, Swipe Deck with spaced repetition, My Blindspots, Decision Pre-Flight, the Printable Deck + Cheat Sheets, and the Field Guide e-book. $29.50$59.
