What the science actually says.

What the questionnaires could not predict

In 2017, Samantha Joel, Paul Eastwick and Eli Finkel published the most direct test of questionnaire-based matching to date. Using machine learning on more than one hundred self-reported traits and preferences collected before speed dates, they attempted to predict which two specific people would desire each other when they met. The models could predict who tended to like others, and who tended to be liked, but pair-specific desire, the thing a matching algorithm actually needs, was not predictable from the questionnaires at all.

Stated preferences carry less than they seem to

A separate line of work asks a related question: do the qualities people say they want predict who actually attracts them? A 2014 review and meta-analysis by Eastwick, Luchies, Finkel and Hunt found that stated ideals predict how people evaluate profiles on paper and fail to predict who draws them in live interaction. A 2020 study by Sparks and colleagues found negligible evidence that people desire partners who uniquely fit their ideals. A registered report published in 2025, run through the Psychological Science Accelerator across 43 countries and 22 languages with more than ten thousand participants, set out to measure how large these effects actually are. The picture across all three is not that stated preferences are meaningless. It is that they carry far less about a specific pairing than the questionnaire model assumes.

What follows from this

Incumbent platforms continue to collect the inputs that are cheap to gather rather than the ones that predict outcomes. Kindred's design takes the evidence at face value: self-report is used only where it is reliable, as a dealbreaker filter; ranking is built from behaviour and reciprocity; and, most importantly, the platform measures real relationship outcomes over time, because the only way to learn what predicts lasting relationships is to observe them forming.

Our commitments

• Outcome data is collected only with explicit, informed consent, separate from the consent to use the platform.
• Findings are published openly, including null results. If behavioural matching does no better than chance, that is a finding, and it will be public.
• Published research uses aggregated and de-identified data. Individual members are never identifiable in anything we release.

For researchers

Kindred is building toward a consented longitudinal dataset that follows introductions through to relationship outcomes at six and twelve months, a resource that, to our knowledge, no commercial platform has ever opened to independent study. We are actively seeking academic partners in relationship science and computational social science, particularly at GTA universities, for co-designed studies and partnership grant applications. If that is your field, we would genuinely like to hear from you: hello@kindredmatch.io.

Join Kindred

References

• Eastwick, P. W., Luchies, L. B., Finkel, E. J., and Hunt, L. L. (2014). The predictive validity of ideal partner preferences: A review and meta-analysis. Psychological Bulletin, 140(3), 623-665. https://doi.org/10.1037/a0032432 (opens in a new tab)
• Eastwick, P. W., Sparks, J., Finkel, E. J., Meza, E. M., and colleagues (2025). A worldwide test of the predictive validity of ideal partner preference-matching. Journal of Personality and Social Psychology, 128(1), 123-146. https://doi.org/10.1037/pspp0000524 (opens in a new tab)
• Joel, S., Eastwick, P. W., and Finkel, E. J. (2017). Is romantic desire predictable? Machine learning applied to initial romantic attraction. Psychological Science, 28(10), 1478-1489. https://doi.org/10.1177/0956797617714580 (opens in a new tab)
• Sparks, J., Daly, C., Wilkey, B. M., Molden, D. C., Finkel, E. J., and Eastwick, P. W. (2020). Negligible evidence that people desire partners who uniquely fit their ideals. Journal of Experimental Social Psychology, 90, 103968. https://doi.org/10.1016/j.jesp.2020.103968 (opens in a new tab)