Can Desire or Love Be Predicted?

Data versus the real world: what compatibility algorithms can’t tell you.

This is DATA, a monthly column by Angela Chen on numbers, nerdery, and what it means to live an evidence-based life.

The summer I was twenty-one, I received the first prophecy I have ever trusted.

I was in Maryland, but the prediction came over the phone, from someone who lived a thousand miles away in the town that is the birthplace of the atomic bomb. I knew Ann was a tarot reader and a scholar of James Joyce. She didn’t even know I existed. I’m a friend of a friend, I explained, cradling the phone to my ear. You know Mark, right?

She did, though not well, and I told her what I knew: After talking with Ann that morning, Mark had been inspired to propose to his ex-girlfriend, a woman he broke up with a year ago and who was now in a new relationship. The idea seemed foolish, but of course I would not have been on the phone were I not also worried for myself. Mark was once so vividly central to my life that later, when we lost touch, I would marvel at how easy it was to forget about him. We never lived in the same place, and I never truly believed we would be more than friends, but the prospect of possibility lost was still enough to compel me to take action—even ill-advised action, like contacting strangers whose information I found by Googling.

Memory can be merciful, and shame has blocked out most of my conversation with Ann. I do remember apologizing over and over. I remember that Ann was sympathetic, and it was more than I thought I deserved. She offered to read my tarot right then, over the phone, the first reading I had ever gotten. The prediction she presented was unlike any I’d heard before or have since.

Most readings of any type are deliberately ambiguous—they are not what the philosopher of science Karl Popper calls “falsifiable,” because they cannot be disproven. Horoscopes and fortune-tellers tend to use vague, general statements that would sound just as plausible if they were about anyone else.

But Ann’s words were different. I would meet my soulmate soon, she said. He would be 5’8” and blond, with blue or gray eyes, and I would meet him in either two weeks or two months.

*

Tarot is not my preferred method of assessing romantic compatibility, but the one I did trust—OkCupid algorithms—had vetted Mark, too. Mark and I were a 99 percent match. When we met, he was in Oklahoma and I was in California, and he was the first of several long-distance friends I made on the site, which I joined junior year of college.

I would use OkCupid for actual dates, of course, but sometimes I would set my location search parameters to “anywhere” and look for people with interesting profiles and high match percentages all over the world. To my mind, match percentages were to be taken seriously, especially given that over the years I had spent enough bored hours on the site to have answered over two thousand questions. It seemed perfectly rational that compatibility was based on shared values and preferences, and to have all that data neatly collected and analyzed and presented seemed nothing short of wondrous. The match percentage was also a useful weapon against touchy-feely friends who claimed that everyone was “great” and “deserved a chance.” In the face of such sunny arguments, I felt obligated to point out that if someone were 90 percent “enemy” and believed that interracial relationships were bad, we probably wouldn’t get along.

Data science had already brought us so many insights in business and behavior; of course we could turn it to romance. And others have taken this line of thought much further: One man “hacked” OkCupid to find love. My friend Jacob did a version of the same thing. For him, it worked so well that he ended up crunching numbers in a spreadsheet to choose between two women he met on the site—and eventually marrying the one the spreadsheet deemed superior.

*

Mark had—has—dark hair. The person after him was my first boyfriend, and the first person I loved. I met him two months later, but he was tall, and he, too, had dark hair.

I forgot about Ann entirely until one day more than two years later, when I found myself describing my then-boyfriend Cory to someone. Blond, I said. Gray eyes, about 5’8”.

Something about the routine rattling off of physical attributes jogged my memory, so I went to check. Cory and I had known each other for over a year by the time we started dating, but we had originally met on OkCupid as well, the summer I was twenty-one. I went back through old emails to try to piece together a timeline. It was true: He had first messaged me two weeks after I spoke to Ann.

*

Stories about love being a numbers game are compelling, but the research is not as positive. In one recent study from the journal Psychological Science, a matching algorithm failed at predicting who would get along after four-minute speed dates. The participants had answered the usual questions—relationship goals, personality traits, preferences—but there wasn’t a straight line between their answers and their desires. Other studies also suggest that matching algorithms simply do not work. Dating sites build their models around principles that are “much less important to relationship well-being than has long been assumed,” according to the authors.

Perhaps we don’t have a good idea of what we want. Even if we did, people regard dating questionnaires with varying degrees of seriousness, and it’s easy enough for a few sloppy answers to greatly change the numerical outcome. Algorithms do not capture our histories, present life circumstances, and how people change over time. The algorithms present theoretical attraction frozen in a vacuum and we are always interacting in the real world, one that constantly shifts.

When it comes to the story of the man who “hacked” OkCupid, the woman he eventually married was a 91 percent match, not 99—and she messaged him first. Next year, I will be a bridesmaid in the wedding of a dear college friend; she and her fiancé were less than an 85 percent match.

And Cory and I broke up, too.

*

Mark and I are still a 99 percent match. The same is true of Cory and I. Mathematically speaking, I should have been equally compatible with both men, but that wasn’t true. The algorithms don’t work via the transitive property. Cory and Mark are very different, opposites in ways, and the things that made me a good match with one weren’t the same that made me a good match with the other.

Cory was careful and reserved, while Mark tended to talk in sweeping generalizations, and I liked Cory for his nuance just as I liked Mark for his sometimes-naive enthusiasm. We’ve all changed since I first met them, yet all the things that matched us are still there. But having known both of them in the real world, the numbers don’t matter anymore. There’s far more important information at my disposal now, information I learned both about them and about myself.

Preferences and values play a role in compatibility, and algorithms do measure something. But a good match percentage is only necessary, not sufficient. When my current partner John messaged me on OkCupid, that same number next to his photo, he was already familiar to me. John didn’t remember, but two years ago to the week, we had chatted on OkCupid before I disabled my account to be with Cory. Stretching back even further, six years ago when I was in San Diego and he Boston, I had viewed his profile and considered sending a message. I suspected we would get along, but decided against it because one particular phrase in his profile annoyed me. The algorithm had served us up to each other for good reason, but it took all those tries for us to meet, for something to happen.

Low matches are very likely to fail, but high matches can as well—or they can take six years to come together. A high match creates a floor on how sour a date, or a relationship, can turn. Similarly, if both people are kind and good at communicating there is typically a limit on how poorly something can turn out. But that information is about limiting the bad. You need much more to know how good it can get.

*

In spite of myself, I have always been interested in what friends call “weird mysticism.” I have dabbled in tarot, buying my own deck and studying the cards, even researching how long each prediction is supposed to last, searching for its statute of limitations. Over time, tarot’s approach of saying nothing definitive seemed like less of a bug and more like a feature. Unlike a percentage, tarot doesn’t provide any semblance of certainty. Instead of focusing my mind on a single outcome, a yes-or-no question, it gives suggestions of what to think about. It can’t tell me whether a door will open or close, but it can point out the shape of the door and the window nearby, offering an ambiguity attractive on its own merits. It provides the opportunity to think about the “more,” a layer above the numbers.

I used to Skype with a tarot reader who lived in Berlin. Unlike Ann, she didn’t say anything that could be cleanly disproven, just offered examples and book suggestions. One April, I requested a reading because I was frustrated with my career and found myself considering the same points over and over. After putting the cards into the Celtic cross, she looked at the position of one of the cards, the King of Cups.

We’ll get to career, she said. But more important is that I think you’ll meet a man who will be very important in an emotional way.

This update did not interest me. I knew the King of Cups typically represented an older man, so was it possible, I asked, that the card represented an older man—a powerful media mogul, perhaps?—who was about to offer me a job?

Not a chance, she said.

Twenty minutes after we hung up, John’s message appeared in my inbox.