Case Study 42.2: Prediction Failure — Why Algorithms Cannot Find Your Person

The Promise

In 2012, a major US-based dating company released a press statement claiming that its proprietary matching algorithm had produced over a million marriages. The algorithm, it said, analyzed 29 dimensions of compatibility, used "scientific principles" to identify optimal partner matches, and had been validated by a team of research psychologists. The company's founder said in an interview: "We've essentially solved the problem of finding the right person."

That same year, Finkel, Eastwick, Karney, Reis, and Sprecher published their comprehensive review in Psychological Science in the Public Interest. Their conclusion: there was no credible scientific evidence that any matching algorithm produced better relationship outcomes than chance would predict beyond a very low threshold.

The Gap

The company was not lying, exactly. They had a million marriages. They had research psychologists on staff. Their 29 dimensions of compatibility were all real constructs with genuine predictive validity for something. The gap between their claim and the scientific assessment is not a gap of honesty but of understanding — specifically, understanding of what kind of thing compatibility is.

What the 29 dimensions can do: Each of the dimensions used by matching companies — agreeableness, conscientiousness, attachment style, values, interests, physical type preferences — has some predictive validity for some outcomes. Attachment security, for instance, is one of the most robust predictors of relationship quality and stability (Simpson et al., 2007). Value similarity predicts long-term satisfaction better than personality similarity in most studies. Physical attraction predicts interest in meeting. None of this is nothing.

What the 29 dimensions cannot do: They cannot predict the quality of the specific dyadic encounter between two people. When Eastwick, Luchies, Finkel, and Hunt (2014) reviewed the literature on individual characteristics and attraction, they found a striking pattern: traits that predicted pre-encounter attraction (the kind of person you think you'll be attracted to) consistently failed to predict actual attraction to a specific person in an actual encounter. The correlation between stated type preferences and actual attraction to real people, in study after study, was near zero.

This means that the 29-dimensional profile of person X cannot predict whether person X will feel attracted to person Y, even if Y matches X's stated preferences closely. The profile captures the map; the encounter is the territory; and the map does not accurately predict the territory.

The Deeper Issue: Compatibility as Emergent

The Finkel et al. (2012) review identified what they called the "missing dyadic chemistry" problem. Most compatibility research examines individuals — their traits, preferences, attitudes. But compatibility is not a property of individuals; it is a property of specific dyadic combinations in specific contexts. The question "are person X and person Y compatible?" is not the same as the question "what traits does person X have + what traits does person Y have?" It is a question about what happens when they interact — and that cannot be derived from the profile data.

A useful analogy: knowing the caloric content, protein, fat, and carbohydrate composition of all the ingredients of a dish does not tell you whether the dish will taste good. Flavor is an emergent property of specific combinations in specific preparations; it cannot be predicted by summing nutritional profiles.

What Algorithms Have Actually Learned to Do (Somewhat) Well

The more honest description of what current matching algorithms can do is: filter out highly incompatible pairs. An algorithm that shows you people in your geographic area with broadly similar values and stated relationship goals, and removes profiles that have clearly incompatible lifestyle preferences, is genuinely useful. It reduces the search space. It increases the probability of a decent first conversation.

That is not the same as predicting who you will fall for. It is more like clearing the field of obvious obstacles before you run through it.

The implications for app design: If compatibility is emergent and not predictable from individual profiles, the most valuable thing a dating platform can do is not better matching but better facilitation of the encounters in which chemistry can emerge. Platforms that enable video interaction, shared activities, voice-first design — that prioritize getting people into actual interaction rather than more sophisticated pre-encounter filtering — are, in principle, better aligned with what the science suggests compatibility requires.

What This Tells Us About Love

The prediction failure is, at a certain angle, a deeply hopeful finding. If compatibility were perfectly predictable from individual profiles, love would be a kind of computation — an optimization problem with a best-fit solution that an algorithm could find. The fact that it is not predictable suggests that something real and important about human connection cannot be reduced to the sum of individual characteristics. The specific encounter — the first conversation, the shared moment, the particular quality of being with this person at this time — matters in ways that no profile, however detailed, can anticipate.

That is not a limitation of science. It is a description of what connection is.

Discussion Questions

1. The case ends by framing prediction failure as "hopeful." Do you find this framing convincing, or is it motivated reasoning — a way of making a limitation sound like a feature? What would need to be true for the failure of prediction to be genuinely good news?

2. If you were building a dating platform based on the science reviewed in this course, what would you design differently from current platforms? What features would you remove and what would you add?

3. The Finkel et al. finding that stated type preferences fail to predict actual attraction is one of the most consistently replicated findings in the attraction literature. Yet most people continue to use type preferences as the primary basis for online dating filtering. What psychological factors might explain this gap between scientific finding and behavior?