In 2020, 86 researchers pooled 43 separate longitudinal datasets, more than 11,000 couples in total, and handed the combined data to a machine learning model with one job: find out which of over 2,000 measured variables actually predict how satisfied two people will be in their relationship. The team, led by Samantha Joel and Paul Eastwick, threw in everything: sexual satisfaction, conflict, life satisfaction, negative affect, depression, attachment style, and dozens of other traits and dynamics researchers have spent decades arguing over. The results, published in the Proceedings of the National Academy of Sciences, are a fairly direct answer to whether love can be predicted in advance, and the answer is not encouraging for anyone hoping a formula exists.
This is one large-scale analysis, not a single simple experiment, and it does not claim to have found every variable that could ever matter. But its scale is exactly what makes the result worth taking seriously: it is not one study finding a weak effect, but 86 researchers combining their best available data specifically to see how strong the strongest possible prediction could be. The project was explicitly designed as a stress test of the field’s existing theories, pooling data collected independently by dozens of separate research teams over many years rather than a single lab’s fresh sample, which makes it considerably harder to dismiss as a fluke of one dataset or one measurement approach.
What actually predicted relationship quality, and by how much
Individual traits, the kind a personality quiz or a dating app profile might capture, such as life satisfaction, negative affect, depression, and attachment anxiety or avoidance, together explained at most about 21 percent of the variation in how satisfied people were with their relationships. Traits about the relationship itself, not the people in it individually, performed better: perceived partner commitment, appreciation, sexual satisfaction, perceived partner satisfaction, and low conflict together accounted for up to 45 percent of the variation. Once those relationship-specific factors were accounted for, adding the individual traits back in did not meaningfully improve the model’s predictions.
The detail that matters most for anyone hoping this points toward a formula is what happened next. Even the relationship-specific variables, the strongest predictors the researchers found, were not useful for predicting whether a couple’s satisfaction would rise or fall over time. They described the relationship as it stood at one moment reasonably well. They did not forecast where it was headed.
Why this is a genuine problem for algorithmic matchmaking
This result lands differently depending on what you are trying to build. A dating platform’s entire pitch typically rests on the idea that enough individual data, personality traits, preferences, demographics, behavioral signals, can be used to predict who will be compatible with whom before the two people ever meet. The Joel and Eastwick analysis directly tested something adjacent to that premise using far more data than any single dating platform’s internal model is likely to have access to, and individual-level traits came in as the weaker half of the equation. The stronger predictors, the ones about how the two specific people interact once they are actually together, cannot be measured before a relationship exists. They are produced by the relationship itself, not by anything knowable about either person beforehand.
That is a structurally different kind of prediction problem than most recommendation systems are built to solve. A system that recommends a film or a product is inferring from a large base of prior individual behavior, and it can be validated by whether people click. A system claiming to predict romantic compatibility is trying to forecast an emergent property of two specific people’s future interaction, one the largest academic effort to date found could not be reliably predicted even after the relationship had already started and researchers could observe how it was actually going.
What “45 percent” actually leaves unexplained
It is worth sitting with what that 45 percent figure means in practice, since it is easy to round it up to “the researchers mostly solved it.” Even taking the strongest set of predictors the entire analysis could find, appreciation, commitment, sexual satisfaction, and low conflict, more than half of the variation in how satisfied people are with their relationships remained unaccounted for. That is the ceiling researchers hit using more combined data than has been assembled on this question before or since. A dating platform working from a single onboarding questionnaire and a handful of behavioral signals is working with a small fraction of the information this analysis had, predicting an outcome that even the full dataset could only partly explain.
What this does not mean for the products themselves
None of this means matchmaking algorithms are useless as a way of filtering an otherwise unmanageable pool of strangers down to a smaller, more plausible set of people, which is a genuinely different and more modest task than predicting long-term compatibility. Reducing thousands of profiles to a shortlist based on shared individual traits and stated preferences is a search problem an algorithm can meaningfully help with. Claiming that same algorithm can forecast whether a specific match will be happy together five years later is a different claim entirely, and it is the specific claim this research does not support.
What this does not prove
The couples in the underlying datasets were drawn from Western countries, including the United States, Canada, Switzerland, New Zealand, the Netherlands, and Israel, and the researchers themselves note that the same patterns have not yet been tested as thoroughly outside those cultural contexts. The study also measured self-reported satisfaction rather than more distal outcomes such as breakup or divorce, and it excluded some variables, including love and trust themselves, as potential stand-ins for the outcome being measured rather than genuine predictors of it. None of this is proof that love is random or that no relationship pattern is ever knowable in advance. What it does show, using more combined data than any single study or startup is likely to gather on its own, is that the specific dynamic between two people accounts for far more of how a relationship goes than anything measurable about either person walking in, and that even that dynamic does not reliably tell you what comes next.
