How it works, and how sure it is.

The first version of this project, a Streamlit app, reported a “dominant” trait, and gave nearly everyone the same one. Replaying its code on held-out respondents, it named Openness for 89% of them. Four problems stacked up:

Training data is the Open Psychometrics IPIP-FFM dataset: anonymous responses to the same 50 statements, collected online between 2016 and 2018 and released for research. Responses were cleaned before use:

The ground truth is standard IPIP scoring: each trait is the mean of its ten statements on the 1–5 scale, with reverse-keyed statements counted as 6 − answer. Percentiles compare that score with the training sample (1,001 stored quantiles per trait, interpolated linearly).

Question order. Starting from nothing, the training code repeatedly adds the statement that most improves the average R² of a linear model predicting all five full-scale scores. That greedy order is fixed for everyone. Under a linear-Gaussian model, how much uncertainty is left after a set of statements does not depend on the answers given, only on which statements were asked, so asking adaptively gains nothing over a fixed order. The order itself is chosen greedily: a good approximation, not a guaranteed optimum.

One model per length. For every prefix length k from 1 to 49 there is a separate linear model per trait: score = intercept + Σ weight × answer, clipped to 1–5. Answers to other traits' statements are allowed to help, because traits correlate a little. At k = 50 the standard scoring formula is used directly.

Explanations are exact. The “what drove this” list shows weight × (your answer − average answer) for each statement. For a linear model these add up exactly to the difference between your score and an average respondent's.

Uncertainty. The “typical error” is the residual standard deviation of the short-form score on the held-out test set, for that trait and that number of statements. The band drawn around each result is narrower or wider depending on where your estimate lands: for every length and trait, training respondents are grouped by their predicted percentile, and the band spans the middle 68% of how far their full-test scores fell from the prediction. On the held-out test set, 69% of people land inside it on the full test. A single pooled band would be the wrong width near the ends of the scale, where scores cannot go above 5 and a fixed number of points covers many more percentiles at the top than at the bottom.

Wording. A result reads “Strongly” when the estimate is beyond the 85th (or 15th) percentile and even the cautious end of its band is beyond the 80th (20th); “Clearly” needs the 70th and 65th (30th and 35th); anything else is “Leaning”. At 20 statements, people told “Strongly” are on that side of the 70th (30th) percentile on the full test at least 97% of the time.

All figures are measured on the held-out test set: R² is the share of variance in the full 50-statement score that the short form explains, RMSE its typical error in points on the 1–5 scale.

A fair test should not favour one trait. The results page names your most distinctive trait, the one whose percentile is furthest from 50. Across held-out respondents answering 25 statements, each trait gets that pick for between 19% and 21% of them, close to an even 20%.

Simply taking the highest percentile would not be even: it lands on each trait for between 17% and 25% of people, with Neuroticism on top most often. That is not a model bias, and not a matter of the score distributions having different shapes (after the percentile transform every trait is close to uniform). It comes from how the traits correlate. Neuroticism runs opposite to Extraversion and Conscientiousness (r = −0.22 and −0.24) and is barely related to the other two, so it has the most negative ties to the rest: when the others are low, Neuroticism is often the highest. Agreeableness goes with Extraversion (r = 0.31), so the two tend to be high together and split the wins. Exact scoring of all 50 statements gives the same pattern (Neuroticism highest for 25%).

How often does the short form pick the same trait as the full 50-statement test? For the highest percentile at 25 statements, 80%. For the most distinctive trait, which is what you see, it depends on how many statements you answer:

The 50 statements are the International Personality Item Pool Big-Five Factor Markers, in the public domain. Goldberg, L. R. (1992). The development of markers for the Big-Five factor structure. Psychological Assessment, 4(1), 26–42. doi:10.1037/1040-3590.4.1.26.

Response data: Open Psychometrics, Big Five (IPIP-FFM) responses, 2016–2018 (download). Thank you to everyone who contributed their answers.

Privacy. The site is static files. Scoring runs in your browser; no answers, results or identifiers are sent anywhere, there are no cookies and no analytics. Share links carry answers after the #, which browsers do not send to servers.

Model v2.2.0, trained 2026-10-10. Take the test.