Two speed-dating studies tested whether self-reported traits and preferences collected before contact could predict romantic desire. They predicted how much a person tended to desire others, and how much a person tended to be desired. They did not predict the distinctive desire of one participant for one particular partner. Two adjacent projects, at later relationship stages, found that reports tied to a particular person carried more information than reports describing a participant in general. Read together, these results point at a measurement boundary rather than a ceiling on prediction: the informative class of report comes into existence only once two people have interacted. Kindred Matchmaking Society is open to members in the Greater Toronto Area. What follows sets out what our design does with pre-meeting information, what it refuses to claim, and the prospective dataset we intend to build with a research partner.

Joel, Eastwick and Finkel opened their 2017 study by naming a proposition that matchmaking practice and close-relationships theory have both entertained: that initial attraction is, to some extent, a product of two people's self-reported traits and preferences (Joel, Eastwick, & Finkel, 2017). Their question was whether reports collected before two people met could predict initial romantic desire.

They ran two speed-dating studies with romantically unattached participants. Before the event, each participant completed more than 100 measures of traits and preferences that earlier researchers had identified as relevant to mate selection. Each then met every opposite-sex participant at the event for a four-minute date. The analysis separated three components of the resulting desire: a participant's general tendency to desire others, a participant's general tendency to be desired, and the desire one participant felt for one particular partner over and above those two tendencies. Machine learning recovered the first two components from the pre-date reports. It did not recover the third.

That decomposition is what makes the result specific rather than merely deflationary. A model can identify who tends to find many people appealing, and who tends to be found appealing, and leave entirely open which two of them will find each other appealing in particular. Pair-specific prediction lives entirely in the third component. A platform that claims to know who suits whom is claiming that component, and the third component is the one the pre-date reports did not reach.

A reviewer will object that this is one measurement set failing in two samples rather than a general limit, and the objection is correct as far as it goes. The studies examined initial attraction after four-minute opposite-sex speed dates, using the reports the researchers chose to collect. Same-sex attraction, longer acquaintance, other populations and other predictors sat outside the test. The warranted conclusion is narrow, and we hold it narrowly: in those samples, the pre-date reports carried information about general tendencies and none about pair-specific desire.

Two later projects examined related questions at later stages. The 2020 project pooled 43 dyadic longitudinal datasets from 29 laboratories, covering 11,196 established couples, and used random forests to compare classes of self-report predictor of relationship quality (Joel et al., 2020). Relationship-specific variables predicted up to 45 percent of variance at baseline and up to 18 percent at the end of each study, with perceived-partner commitment, appreciation, sexual satisfaction, perceived-partner satisfaction and conflict leading that class. The leading individual-difference predictors were life satisfaction, negative affect, depression, attachment avoidance and attachment anxiety. Change in relationship quality over the course of a study was largely unpredictable from the baseline variables examined. The 2023 preregistered study followed reports about specific potential partners during early relationship development, and found that adding individual-difference reports did not reliably contribute beyond target-specific reports when predicting romantic interest (Eastwick et al., 2023).

Neither project repeats the 2017 test, and we do not present them as replicating it. One concerns relationship quality among couples already together. The other concerns interest in partners a participant has already identified. Their value here is that they arrive at the same distinction from two different directions.

The information that matters arrives only after contact

Across all three projects, the reports that carried information were tied to a particular person or a particular relationship: what this person thinks about that one. The reports that carried less described a participant in general, without reference to anyone.

That distinction has a consequence for a platform, and the consequence is about timing rather than about difficulty. Before two people interact, a platform can collect life facts, individual differences and stated preferences. It cannot collect either person's judgement of the other, because neither judgement exists. No questionnaire, at any length, contains a response about a counterpart nobody has met.

This makes pre-meeting information useful for a different job than the one it is usually sold for. It can establish whether an introduction satisfies the constraints each person has set. Which of those constraints are genuine and which are preferences that loosen once a real person is in front of you is a separate question, taken up in our piece on dating dealbreakers. Producing the informative class of report is the work of the introduction itself. The design question for a platform is therefore one of speed and cost rather than one of squeezing more prediction out of questionnaires: how quickly, and at what price to the people involved, can two people be brought to the point where partner-specific information exists at all.

One boundary runs across everything above, and it concerns who was studied. The 2017 speed-dating events paired opposite-sex participants. A field-level audit of 1,762 studies published in eight mainstream psychology and relationships journals, across 1996 to 2000 and 2016 to 2020, examined reporting and representation on gender, sexual orientation, regional context, socioeconomic status and race (McGorray et al., 2023). It found reporting that remained limited even where practices had improved, and that frequently centred Whiteness, obscured or overlooked sexual-orientation diversity, and treated the United States as the default context. The audit concerns reporting and representation rather than predictor performance, so how these measures behave in another population is an open question. What the audit closes off is the assumption that they behave the same way.

What we built, and what we refuse to claim

Kindred Matchmaking Society is a federal Canadian not-for-profit corporation. It opened to members in the Greater Toronto Area in September 2026, and everything in this section describes the design as built rather than a result. The intended first pool is an older, more diverse, urban Canadian adult population in the Greater Toronto Area.

The design separates three operations that a compatibility score collapses into one: eligibility, ranking and interaction.

Eligibility comes first. We apply twelve member-set dealbreaker dimensions as rules, covering life facts including children, faith, marriage intent, relationship structure, food practice, family involvement, language, pets, alcohol and age. These are constraints rather than weighted preferences: a pair either satisfies them or falls out before ranking. Marriage intent is one dimension among the twelve rather than an assumed endpoint, and cultural, religious and family-structure constraints belong to the people who set them rather than being friction for the system to route around. Gender direction is a rule as well, and it must be mutual: members are matched on their own gender identity, and each member states whether they want to be matched with men, women, non-binary people, any combination of those, or everyone. Geography is handled by pooling members within an active market rather than by treating distance as another preference. Everyone you can be matched with has verified their identity. Each member holds at most three concurrent introductions.

Prior introductions and blocks remove a pair from consideration. The procedure produces no compatibility score and orders no pair by predicted relationship success.

One measured dimension supplies the ordering. We administer the Experiences in Close Relationships Scale, Short Form, a 12-item self-report instrument developed across six undergraduate samples at one public university whose participants were predominantly White, yielding two continuous dimensions: attachment-related anxiety and attachment-related avoidance (Wei et al., 2007). Only avoidance is used. For each eligible pair we add the two continuous avoidance scores and multiply the sum by a fixed weight of 0.1. Every eligible pair in a market is ordered by that penalty, lowest first, and taken greedily subject to each member's limit of three concurrent introductions. The result is a stable matching in the technical sense, because every member's ordering over partners is induced by the same pair ordering. There is no cutoff, avoidance excludes no pair, anxiety does not affect ranking, no derived security score is used anywhere in matching, and the matcher does not read photographs.

One consequence of that specification is worth stating rather than leaving to be found. At launch the avoidance penalty is the only term in the ordering, so the constant multiplying it does not change the resulting order at all. The weight exists so that this term can be balanced against others once others are added, and describing it as low weight would be misleading about what it currently does.

The strongest objection to the rule is that nothing cited here warrants it. The 2020 project found attachment avoidance and attachment anxiety among the leading individual-difference predictors of relationship quality, in established couples, on a different outcome from the one we are ordering for. That is a reason to record the dimension and test it. It is not a reason to act on it. The instrument's own paper reports measurement properties in its own samples and did not examine matching outcomes. We treat the avoidance penalty as a hypothesis with a named and insufficient warrant. What makes it cheap to be wrong about is not the size of the constant: it is that the term reorders an already-eligible set, excludes nobody, and is written into the match record in full, so that it can be evaluated against outcomes and retired.

Two questions about the instrument are open, and we would rather state them than be asked. Whether the ECR-S retains its measurement properties across our intended population is unknown. Permission to administer it was requested from its authors in July 2026 and was not granted, and we proceeded without it, treating that as an accepted exposure rather than a settled matter.

The interaction is where the design does its actual work. Each introduction runs four ordered prompted rounds: a light opener, everyday life and habits, values and relationships, and a reflective or future-looking round. Each person answers a round privately, and both answers reveal together once both are in. Once the first pair of answers has revealed, free-text conversation opens in one continuous thread and stays open for as long as the introduction lasts. Photographs stay hidden until 48 hours have passed since the introduction began and all four rounds have revealed, and release is then independent and blind: neither person is told the other's choice, and photographs become visible only if both release. A display name is visible throughout, and verified legal names are never shown to another member. Either person may end an introduction at any time.

The purpose of that sequence follows from the measurement boundary rather than from any claim about attraction. It exists to produce partner-specific information early, in a form both people can act on, and to do so before appearance dominates the judgement. Whether a slow or sparkless first meeting tells you much on its own is a related question, taken up in a separate piece on first dates. Whether it changes any recorded outcome is untested, and it is among the things the dataset is meant to find out.

The dataset, and the questions it can reach

For every introduction we will record the factors the matcher used: both avoidance scores, the weight, the penalty actually applied, the gender-compatibility result, the dealbreaker outcome and the rank. Outcomes will be recorded as a progression ladder: conversation participation, mutual photo release, contact exchange, first date, relationship progression and successful exit from the platform.

That combination is the part of this programme we think is worth a researcher's time, and it is rare. There are datasets describing couples once they are couples. There are datasets describing attraction in a laboratory. A prospective record linking the exact factors used to select an introduction to what became of that introduction, in a general adult population outside a university, is something we have not found in the published literature. If that is wrong, we would like to be told, because it would change what this programme is for.

The programme's limits are structural, and we would rather set them out than have them found. The data will be observational. Introductions are not randomly assigned, and the ranking rule itself determines which pairs are exposed to an introduction, so any association we observe carries the imprint of the rule that produced it. Members select themselves into the platform. The pilot covers one metropolitan market. Outcomes at the far end of the ladder, durable relationships in particular, will arrive slowly and stay sparse for a long time. Together these conditions rule out causal claims about the ranking weight, the prompted sequence or the delayed photo release. A design capable of supporting such claims would look different from this one, and building one is a conversation we are open to having.

The first question we can answer is bounded: does avoidance add prospective information about movement through the recorded ladder once the other match-time factors are accounted for, and where does movement stop? Answering it needs complete match-time records, the penalty actually applied, and outcomes for the whole set of introductions rather than for the successful ones.

Kindred Matchmaking Society intends to work with researchers on evaluation design, data governance and publication, and the programme is built to publish null results as readily as positive ones. Mitacs Accelerate and SSHRC Partnership Engage are both open to a not-for-profit partner organisation, and neither requires that funding flow to Kindred Matchmaking Society. We would rather define the evaluation before the dataset exists than after it. Correspondence to Bilal Ahmad, Founder and Executive Director, at bilal@kindredmatch.io.

Keep reading

References

  • Eastwick, P. W., Joel, S., Carswell, K. L., Molden, D. C., Finkel, E. J., & Blozis, S. A. (2023). Predicting romantic interest during early relationship development: A preregistered investigation using machine learning. European Journal of Personality, 37(3), 276-312. https://doi.org/10.1177/08902070221085877
  • Joel, S., Eastwick, P. W., Allison, C. J., Arriaga, X. B., Baker, Z. G., Bar-Kalifa, E., Bergeron, S., Birnbaum, G. E., Brock, R. L., Brumbaugh, C. C., Carmichael, C. L., Chen, S., Clarke, J., Cobb, R. J., Coolsen, M. K., Davis, J., de Jong, D. C., Debrot, A., DeHaas, E. C., ... Wolf, S. (2020). Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies. Proceedings of the National Academy of Sciences, 117(32), 19061-19071. https://doi.org/10.1073/pnas.1917036117
  • Joel, S., Eastwick, P. W., & 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
  • McGorray, E. L., Emery, L. F., Garr-Schultz, A., & Finkel, E. J. (2023). "Mostly White, heterosexual couples": Examining demographic diversity and reporting practices in relationship science research samples. Journal of Personality and Social Psychology, 125(2), 316-344. https://doi.org/10.1037/pspi0000417
  • Wei, M., Russell, D. W., Mallinckrodt, B., & Vogel, D. L. (2007). The Experiences in Close Relationship Scale (ECR)-Short Form: Reliability, validity, and factor structure. Journal of Personality Assessment, 88(2), 187-204. https://doi.org/10.1080/00223890701268041