Published 2026-09-26 · 6 min read

What Actually Predicts a Happy Relationship?

Eighty-six researchers pooled data from 43 long-running studies of couples and used machine learning to find out what really predicts relationship satisfaction. The answer was not what most compatibility quizzes assume.


In 2020, a collaboration of 86 relationship researchers did something almost nobody in the field had done before: they pooled their raw data. Instead of one lab running one study on a few hundred couples, the team led by Samantha Joel and Paul Eastwick combined 43 long-running studies of romantic couples into a single dataset, then used machine learning to test hundreds of variables at once - personality traits, demographics, sexual satisfaction, attachment style, how partners communicated, and more - to see which ones actually predicted how satisfied people were in their relationships.

The project, published in the Proceedings of the National Academy of Sciences, was partly a response to a problem the field already knew it had. Countless studies had found this or that trait "predicts" relationship satisfaction, but each study tested only a handful of variables on one sample, which makes it easy to overstate how much any single factor matters once you consider everything else going on in a relationship at the same time.

The variables that barely moved the needle

The headline finding surprised a lot of people who build compatibility quizzes for a living: traits belonging to your partner, and static facts about the relationship itself - how long you have been together, whether you live together, income, even most personality traits measured in isolation - explained very little of the variation in how satisfied people said they were. A partner's neuroticism or your shared interests mattered far less than popular relationship advice implies.

What actually showed up

What did carry weight was how a person perceived the relationship they were currently in - how satisfied they felt, how committed they felt, and above all how responsive and available their partner seemed to be day to day, independent of what that partner might have said about themselves on a personality inventory. In other words, the strongest predictors were not facts you could look up about your partner from the outside. They were the felt, ongoing experience of the relationship from the inside, measured in the present rather than inferred from a checklist filled out before you ever met.

It is worth being honest about the limits of even this large study: the models, while better than chance, still left most of the variation in relationship satisfaction unexplained. The authors were careful to frame this as a finding in itself - close relationships are shaped by so many small, situational, and idiosyncratic events between two specific people that a small set of general variables, however well measured, was never going to fully predict how any one couple would feel.

Gottman's lab and its limits

No discussion of relationship prediction avoids John Gottman, whose lab spent decades filming couples having short conversations about a point of conflict, coding their facial expressions and tone, and even measuring heart rate and other physiological signals during the discussion. In a landmark 1992 paper with Robert Levenson, couples whose conflict conversations were marked by contempt, defensiveness, and a pattern of one partner withdrawing while physiologically "flooded" were more likely to have separated when the researchers followed up years later.

That is a genuine and interesting finding, but the extraordinary predictive-accuracy figures attached to Gottman's name in popular coverage go well beyond what any single peer-reviewed study demonstrated on an independent sample. The couples in these studies were relatively small in number, self-selected to be filmed discussing conflict in a laboratory, and the follow-up classifications were built and checked mostly within the same research program rather than confirmed by an outside team running a genuinely blind prospective test at that scale. That does not make the underlying observations wrong - contempt and stonewalling are worth taking seriously in your own relationship - but it does mean the specific percentages often quoted deserve more scepticism than they usually get.

A related 1999 paper by the same two researchers looked at what happened after a conflict conversation ended rather than during it: how quickly a couple's mood and physiology settled back down once the disagreement was over. Couples who recovered - or "rebounded" - faster tended to fare better later, which fits a theme that keeps reappearing in this area: it is rarely whether a couple disagrees that matters most, since every couple does, but what the disagreement does to them afterward, and how quickly the relationship absorbs it.

What this means for you

Put the two lines of research together and a fairly consistent, if less dramatic, picture emerges: the outside facts people obsess over when sizing up compatibility - shared hobbies, a partner's job, how a first date went, whether your personalities are supposedly "opposite" or "the same" - carry less predictive weight than how the relationship actually feels to be in, day to day, right now. Do you feel your partner notices when something is wrong and responds to it? Do you feel satisfied more often than not? Those ongoing, present-moment perceptions are closer to the center of what the data can actually support.

If you want to look at your own relationship through that lens rather than a generic checklist, our relationship satisfaction test and partner responsiveness test measure exactly those present-moment perceptions, and our relationship flags test looks at some of the conflict patterns that Gottman's work suggests are worth paying attention to. None of these tools can predict your specific future - no test, and honestly no researcher, currently can - but they can help you notice what you are actually experiencing, which the research suggests matters more than most people assume.

Sources

  • Joel, S., Eastwick, P. W., Allison, C. J., et al. (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. doi:10.1073/pnas.1917036117
  • Gottman, J. M., & Levenson, R. W. (1992). Marital processes predictive of later dissolution: Behavior, physiology, and health. Journal of Personality and Social Psychology, 63(2), 221-233. doi:10.1037/0022-3514.63.2.221
  • Gottman, J. M., & Levenson, R. W. (1999). Rebound from marital conflict and divorce prediction. Family Process, 38(3), 287-292. doi:10.1111/j.1545-5300.1999.00287.x
Advertisement
Ad space · article