A platform that crowdsources the insights of experts to make predictions on events has come up short in its second attempt to call the Kentucky Derby. It got last year’s race exactly right.
Unanimous A.I., a company touting the power of collective intelligence to provide insights into the future, correctly predicted the top four finishers of the 2016 Derby: Nyquist, Exaggerator, Gun Runner, and Mohaymen. Anyone who bet their prediction of the top four finishers would have scored a so-called “superfecta” that paid out on odds of 540 to 1.
That success earned Unanimous this year an official handicapping partnership with Churchill Downs, the racetrack where the Kentucky Derby is held, and the company once again used its AI platform to analyze input from “some of the best racing minds in the world.”
But the system didn’t turn out to be nearly as accurate this year. Two of its top four picks missed expectations significantly, and it failed to foresee the emergence of one dark horse. (After all, that’s why they call it a dark horse.)
Get Data Sheet, Fortune’s technology newsletter.
This year’s top three picks from Unanimous were Classic Empire (actual finish: 4th), McCraken (actual finish: 8th), and Irish War Cry (actual finish: 10th). The Derby’s ultimate winner, Always Dreaming, was ranked fourth by the predictive system, with only a 65% chance to finish in the top four.
In its post-race analysis, Unanimous points out that this year’s Derby field was “flat and unpredictable,” unlike a 2016 race that had clearer favorites. The biggest outlier this year was a horse called Lookin at Lee, a 30-to-1 longshot that finished second. Not a single expert in the Unanimous pool picked that horse to place.
The company says its swarm analysis still outperformed individual experts, whom averaged 1.6 correct picks for a Top 5 finish (compared to the swarm’s two correct picks).
Still, the company seems to accept that it essentially got lucky with its 2016 picks. “Some outcomes are just not predictable,” it wrote after the race. It’s a lesson that the missteps of much-hyped big data efforts, such as attempts to predict last year’s U.S. election, continue to drive home.
