We can’t tell you who you’ll fall for. Neither can anyone else.
In 2017, researchers at Northwestern and Western Ontario tried. They collected more than a hundred traits and preferences from speed-dating participants — everything from values and personality to what each person said they wanted — and applied machine learning to predict who would be drawn to whom. The models could predict how much a person would generally like others, and how much others would generally like them. What no model could predict, from any combination of what people reported beforehand, was the specific spark between two particular people.
We take that finding seriously, and it shapes everything here. ShidduchAI does not claim to predict chemistry. It is built to do the thing that can be done well: rule out the matches that were never going to work, and surface the ones with real reason behind them. Then it gets two people into a fifteen-minute meeting and gets out of the way.
Joel, Eastwick & Finkel (2017), Psychological Science. Replicated by Eastwick et al. (2023).
What it can tell you.
A great deal, as long as the question is about durability rather than attraction. The strongest and most replicated finding in the literature is also the least surprising one: similarity predicts stability, and complementarity does not. More than two hundred studies since the 1950s point the same way, and shared religious commitment is among the most robust versions of it — both the denomination and, more importantly, the level of observance.
Character and emotional steadiness come next. Emotional instability is the most reliable personality risk marker and conscientiousness the most reliable protective one — though the honest framing is that these effects are small, much smaller than values and religious alignment, and we weight them accordingly. Family-of-origin stability is a stronger and more measurable factor: parental divorce roughly doubles the odds that a marriage ends in divorce.
Then attachment security, measured with a validated twelve-item instrument rather than questions we invented; financial values, where disagreements about money predict divorce more strongly than disagreements about anything else; and age at marriage, which follows a curve rather than a line — the risk falls year by year through the twenties and begins to rise again after about thirty-two.
A community survey of 1,736 Orthodox respondents found good character the single most cited reason marriages succeed, and bad character the most cited reason they end, with similar religious levels close behind. It is an opt-in panel reporting perceptions rather than a probability sample, so we treat it as directional intelligence about this community and lean on the broader literature for weight.
Four stages, in this order.
The order matters as much as the weighting. Nothing is scored until it has cleared what is not negotiable, and nothing is suggested on score alone.
01
Halachic requirements and dealbreakers
Applied first, and absolutely. Questions of status and lineage and the constraints that follow from them, together with whatever either of you has said you will not cross. Anything that fails here is never scored, never ranked, and never shown to anyone.
02
What each of you is looking for
In both directions or not at all. You have to fit what they described, and they have to fit what you described. A pairing that only works one way is not a match, and reciprocal matching is one of the few places where the research is unambiguous about method.
03
Weighted compatibility, across three tiers
What survives is scored — and each factor is weighted by how well the evidence actually supports it, rather than by how interesting it sounds. Where an answer is missing, confidence in the pairing drops. Nothing is ever guessed to fill a gap.
- Tier one
Strong evidence
Religious life and hashkafa · values and life goals · character and emotional steadiness · family-of-origin stability · age at marriage
- Tier two
Moderate evidence
Attachment security · money and financial values · conscientiousness · education and background
- Tier three
Weak, exploratory
Conflict style · narrative maturity · shared interests and lifestyle
04
A final read, and a veto
Every remaining pair is read in full — the structured answers, and the way each of you describes your own life. That read can hold a suggestion back. It cannot force one through.
What we refuse to measure.
Every one of these is available to us, and some of them would be easy to sell. They are not here because they do not survive contact with the evidence.
- Attractiveness
- We do not score anyone’s face. Symmetry algorithms are unreliable, carry demographic bias, and attraction is exactly the thing the research shows is personal and unpredictable. You see photos and a video, and you decide for yourself.
- Astrology
- Birth dates tell us your age. Nothing else. Readings feel accurate because they are written to feel accurate about almost anyone — there is no predictive validity here, and we will not quietly use something we can’t defend.
- Personality types
- Four-letter type systems have no established validity for predicting relationship outcomes. We use measures that do.
- “Shared wounds”
- The idea that people should be matched by their childhood injuries is popular in some therapeutic writing and unsupported by evidence. We ask about your life because your story tells us about your values and your resilience — not to pair you by your pain.
What the shidduch system already got right.
It is worth saying plainly, because it cuts against the usual posture of a technology company arriving in a community: on the questions where the evidence is strongest, the tradition was already there.
Shared religious life and shared values are the best-supported predictors of a marriage that lasts — and they are the first thing anyone asks about here. Family involvement and community embeddedness show up in the literature as protective, not as interference. Meeting with intention, knowing why you are being introduced, having people around you who know both families: these are not quaint. They are the parts that hold up.
What the system costs is search. Finding the small number of people who genuinely fit, across a community larger than any one person’s circle, is slow and depends on who happens to know whom. That is the part worth changing, and it is the only part we are trying to change. The judgment stays where it belongs — with you, your family, and a shadchan if you want one.
Limits, in our own words.
No algorithm can promise a good marriage, and we won’t. What we can promise is that nothing here is guesswork dressed up as science, that we will tell you why we suggested someone, and that every decision is yours.
The weighting described on this page is where the published evidence points today, not a finished answer. As real outcomes accumulate we will refine it against what actually happened in this community rather than what held in largely Western, largely secular research samples — and where a factor turns out not to discriminate, we will drop it.
The framework is not a secret: it is published research, cited below, and you are welcome to read every paper it rests on. The implementation is ours.
References
- 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.
- Eastwick, P. W., Joel, S., Carswell, K. L., Molden, D. C., Finkel, E. J., & Blozis, S. A. (2023). European Journal of Personality, 37(3), 276–312.
- Joel, S., et al. (2020). Machine learning uncovers the most robust self-report predictors of relationship quality. Proceedings of the National Academy of Sciences, 117(32), 19061–19071.
- Amato, P. R., & DeBoer, D. D. (2001). The transmission of marital instability across generations. Journal of Marriage and Family, 63, 1038–1051.
- Wei, M., Russell, D. W., Mallinckrodt, B., & Vogel, D. L. (2007). The Experiences in Close Relationship Scale (ECR)–Short Form. Journal of Personality Assessment, 88(2), 187–204.
- Dew, J., Britt, S., & Huston, S. (2012). Examining the relationship between financial issues and divorce. Family Relations, 61(4), 615–628.
- Stanley, S. M., Amato, P. R., Johnson, C. A., & Markman, H. J. (2006). Premarital education, marital quality, and marital stability. Journal of Family Psychology, 20(1), 117–126.
- Myers, S. M. (2006). Religious homogamy and marital quality. Journal of Marriage and Family.
- Hwang, W. (2021). Religious homogamy and marital outcomes. Family Relations.
- Carroll, J. S., & Doherty, W. J. (2003). Evaluating the effectiveness of premarital prevention programs. Family Relations.
- Kraft, K., & Neimann, S. Effect of labor division between wife and husband on the risk of divorce. IZA Discussion Paper 4491.
- McAdams, D. P., & Guo, J. (2015). Narrating the generative life. Psychological Science.
- Li, T., & Chan, D. K.-S. (2012); Hadden, B. W., et al. (2014); Candel, O.-S., & Turliuc, M. N. (2019). Meta-analyses of adult attachment and relationship satisfaction.
- Meta-analysis of 51 study effects on personality and divorce (N = 25,153). Journal of Divorce & Remarriage (2021).
- Wolfinger, N. H. Age at marriage and the risk of divorce. Institute for Family Studies, analyses of the National Survey of Family Growth (2006–2010, replicated 2011–2013).
- Nishma Research (2025). Survey of the Orthodox Jewish community (N = 1,736, including 351 divorced respondents).
When it is time, we will introduce you.
Begin your profile