by Gina Keating
They next went looking for environmental and psychological factors that affected how and why people rated movies the way they did. Were subscribers more or less generous when they rated on weekends versus weekdays? What effect did rating a lot of movies at one time have? Did people rate movies differently according to their moods, and if they did, how could that be quantified? Did a person’s propensity to be a tough rater or a generous one change over time, and if so, how and why?
Each of those questions became an equation of its own to be tested and, if the results were consistent and relevant, added to a stew of equations that made up their winning formula.
As the improvements to Cinematch piled up in painful halves and tenths of a percent, a small subset of films eluded classification and emerged in the second year as a major barrier between the Netflix Prize contestants and their $1 million payday. These movies generally were of an ironic or polemic nature, and they sharply divided audiences and critics on whether they were masterpieces or crap.
Chief among this group was the quirky independent film Napoleon Dynamite—the title that accounted for the largest error rate in all of BellKor’s models—as well as politically polarizing movies like Fahrenheit 9/11, Michael Moore’s documentary about the terror attacks on New York and Washington and the second Iraq war.
Predicting which side of the ratings gulf subscribers would come down on when rating films like I Heart Huckabees, Lost In Translation, The Life Aquatic with Steve Zissou, and The Passion of the Christ was a crapshoot. There was just no telling from previous ratings how people would feel about these films.
Bell reasoned that the solution to the Napoleon Dynamite problem lay not only in finding neighbor films but in teaching the algorithm when it did not know enough about a subscriber to make a prediction at all. The result was an equation that discounted a subscriber whose ratings were too paltry or who gave too many ratings for one type of film or a meager number of consistently high or low ratings.
Despite seminal insights in the second year, the teams squeezed out just an additional 1 percent improvement over the previous year’s progress. BellKor in Big Chaos collected another fifty-thousand-dollar Progress Prize to add to a growing prize trove that included a kitschy replica of a Hollywood Walk of Fame star that they won the previous year and placed in the lobby at AT&T Shannon Laboratory.
Netflix’s Bennett, who retired in 2009, wondered if the $1 million prize would ever be claimed. The contest started up again in earnest in January; the leaderboard leaped to life as teams struggled to close the gap of less than 1 percent improvement over BellKor in Big Chaos’s results to capture the big prize.
Teams began combining on a grand scale in hopes that marrying their methodologies would help them bridge the last few tenths of a percent and carry them over the 10 percent threshold. BellKor in Big Chaos also went looking for fresh ideas. They found two French Canadian software programmers, Martin Chabbert and Martin Piotte, who were charging up the leaderboard by combining the Progress Prize–winning formula with their own, unorthodox, solutions.
Neither Chabbert nor Piotte, who called themselves Pragmatic Theory, had any formal training in data-mining methods, and they had purposely refrained from studying the research stemming from the prize during its first two years. They preferred, they said, to approach the problem by finding patterns in the data or psychological aspects of the subscribers and translating them into working software models. They rejected external movie data and concentrated on predicting ratings rather than on trying to explain them with their formulas.
“The algorithms that find actual patterns in the data, in infinite shades of gray, are much more powerful than any sort of meta-data, which assigns to black-and-white boxes,” Chabbert said.
Their creativity moved the needle for their combined team, now dubbed BellKor’s Pragmatic Chaos, a crucial 0.65 percent—bringing them over the 10 percent threshold on June 26, 2009.
The Netflix Prize rules called for a thirty-day last call period for competing teams to match BellKor’s presumably winning submission—a period that they all found nerve-racking. Several top-ranked teams combined as The Ensemble, and on July 25, 2009, they submitted a solution that bested BellKor by 0.04 percent.
In the frantic twenty-four hours before the competition closed for good, Koren and Big Chaos were in constant touch, trying to eke out another tenth of a percentage point or two from their combination of equations. They finally turned in their last and best solution and waited in four separate countries for the contest to close. Twenty minutes later it appeared that the Ensemble’s results had bested BellKor by one-hundredth of a percent.
For about an hour after the contest closed, Netflix went dark. Volinsky, at a family vacation in Seattle, sneaked away periodically to check his e-mail. Nothing. When they had won the two Progress Prizes, Netflix had notified them that they had won within minutes.
A dejected Volinsky conferred with Bell back in New Jersey and the other team members and decided to turn off his cell phone. He couldn’t resist hitting the refresh button one more time, and as the e-mails loaded he saw it: a message from Netflix.
They had won.
The members of BellKor’s Pragmatic Chaos met in person for the first time when they picked up the medals Hastings awarded them at a press conference at the Four Seasons hotel in New York City attended by the director of AT&T Laboratory, Hastings, Netflix chief scientist Neil Hunt, members of The Ensemble, and a phalanx of journalists.
Hastings had not wanted to travel to New York for the press conference, preferring to have it in Los Gatos, but Swasey insisted. He knew the prize and the people who had vied for it for nearly three years had captured the attention of the world’s scientific community as well as a decent number of ordinary people. What they had accomplished was worth celebrating with an all-out, formal ceremony with speeches, gold medals bestowed on the winners, and a wide-ranging dialogue with journalists.
After the ceremony, the winning team held a technical briefing to show how they had won. Swasey, who had been thrilled by the turnout, was amazed that nearly all of the media stayed to hear the extremely arcane hour-long talk. The event punctuated three years of steady behind-the-scenes work by Swasey as he ginned up media interest in a science contest. He celebrated the bonanza of headlines he had generated that day by going around the corner from the Four Seasons for an expensive and mediocre sushi dinner alone—feeling slightly forlorn that all the fun was all over.
The format for the Netflix Prize press conference, with its branded banners, eye-catching props, and high-tech hardware, later became the template for the normally low-key Netflix’s barnstorming rollout of its international services.
Bell and Volinsky did not keep the prize money, but each designated a charity where they wanted the funds donated. As required by the contest rules, AT&T granted a license for the winning algorithm to Netflix and applied it to its own U-verse television service to monitor users’ television habits and suggest programming they might like.
The contest resulted in a recommendation system so sophisticated that it could read peoples’ movie tastes from behavioral clues and no longer needed much input from the ratings system—especially when it was paired with a video streaming application. For example, the system could soon determine that a particular subscriber watches comedies on certain weekday evenings, or binges on episodes of a TV cop drama on the weekends, or rewinds when a particular actor or scene appears.
“We’re getting information about what you like without you having to do anything,” Volinksy told me after the contest. Subscribers no longer even have to rate movies, because a program embedded in a set-top box or on the Netflix Web site, monitored what shows and movies they watched and how they watched them to figure out whether the selection was memorable, and how to duplicate the experience with films available in the streaming library. If the algorithm chooses hits more
often than misses, it captures the essential ingredient for a successful brand—our trust.
The Cinematch algorithm represents the marriage of marketing and technology that conferred such extraordinary success on Netflix. Because consumers found what they wanted among a limited DVD library, they left the video store and followed Netflix online. The trust they placed in the company—fostered by Randolph’s intuitive user interface and peerless customer service and coupled with Hastings’s beautiful algorithms—allowed it to smoothly shift the movie rental paradigm to streaming, where so many others had failed.
CHAPTER TWELVE
HIGH NOON
(2007–2008)
A WEEK AFTER THE INSTANT Viewing launch, Hastings and his team arrived at the Sundance Film Festival, where both Netflix and Blockbuster customarily held court with independent filmmakers in snowy Park City. For the first few years after Netflix’s launch, Hastings shut down the company and flew every employee to Park City for the festival. Staffers would go to screenings and hang out at the old church Hastings had converted into a vacation home. It was his way of emphasizing that they were part of a great company that was doing important things.
That Sunday Hastings sat alone in his church-turned-chalet and waited for his rival to arrive. It was a gray and cold January afternoon. The snowy lane outside the modest redbrick meetinghouse echoed with the voices and boots of festival-goers meandering down the hill to the little town’s makeshift screening rooms to watch movies, make deals, and be seen.
In a few hours Hastings would follow them down to the invitation-only party that Netflix sponsored every year in a warehouse fitted with a red carpet for celebrity arrivals, a constellation of open bars, and a dance floor. The party had become rather famous, and the year before, Evangelist and the Blockbuster Online crew had posed as Netflix employees to try to crash it. They had been stopped at the door and turned away when one of the twenty-somethings had claimed to be Hastings.
This year’s party celebrated Little Miss Sunshine, which had debuted a year earlier at Sundance and now was up for four Academy Awards. The Netflix marketing team had decorated the warehouse’s interior to recall the film’s suburban desolation—from the gaily-patterned plastic tablecloths to the buckets of Dinah’s Chicken featured in the film.
The decorators had even parked an ancient yellow VW bus, like the one Greg Kinnear drove in the movie, in one corner. The party was designed to echo the ironic cool that appealed to consumers about Netflix, and that the now ten-year-old company had carefully cultivated.
Sundance was the top marketing event of the year for Netflix partly because the festival’s antiestablishment yet glitzy vibe echoed Netflix’s journey in the entertainment world, from dark horse to suddenly significant. Hastings went all out to get the Netflix logo seen by the right people in the movie industry—studio chiefs and filmmakers whose power to grant streaming rights would be crucial to fulfilling the company’s appeal as the place to go online for movies.
In 2007, Hastings recruited dozens of his young employees and turned them out on the streets of Park City outfitted in red parkas bearing Netflix’s red-and-white logo, like walking billboards. The company’s marketing department sprinkled Netflix beanbag chairs, ball caps, scarves, and other promotional items into swag bags and hospitality suites.
Evangelist took note of the near saturation of Park City’s streets with the red Netflix logo from the rented chalet where he was staying with Craft and Cooper and counterattacked by hiring half a dozen models to strut around in tight jeans and white parkas bearing the Blockbuster Online logo.
Hastings had an urgent and confidential mission. On this trip one of his main concerns was finding the right time to approach Antioco about doing a deal to save Netflix—by making Total Access, or preferably Blockbuster Online, disappear.
• • •
NEARLY A MILLION new subscribers joined Blockbuster Online in the two months after Total Access launched, and market research showed consumer opinion nearly unanimous on one important point—the promotion was better than anything Netflix had to offer. Hastings figured he had three months before public awareness of Total Access began to pull in 100 percent of new online subscribers to Blockbuster Online, and even to lure away some of Neflix’s loyal subscribers.
Hastings had derided Blockbuster Online as “technologically inferior” to Netflix in conversations with Wall Street financial analysts and journalists, and he was right. But the young, hard-driving MBAs running Blockbuster Online from a Dallas warehouse had found the one thing that trumped elegant technology with American consumers—a great bargain.
His momentary and grudging admiration for Antioco for finally figuring out how to use his seven thousand–plus stores to promote Blockbuster Online had turned to panic. The winter holidays, when Netflix normally enjoyed robust growth, turned sour, as Hastings and his executive team—McCarthy, Kilgore, Ross, and chief technology officer Neil Hunt—pondered countermoves.
Models created by Kirincich and Ziegler showed that Blockbuster faced financial ruin if it continued giving away two-for-one rentals for much longer than a quarter or two. In the end, the Netflix team agreed that Hastings had to try to convince Antioco to wind down the promotion before it killed both companies. The Sundance meeting was his chance to do just that.
• • •
THE COMMUNICATIONS REQUIRED to bring the two chief executives together on January 21, 2007, were almost as complex as diplomatic hurdles in a visit between two heads of state. After intermediaries at their companies finally connected them via three-way text messaging, Hastings and Antioco agreed to meet later the same day at Hastings’s chalet.
Neither Antioco nor Shepherd had to ask what their rival wanted to discuss. Since arriving at Sundance a few days earlier, the Blockbuster team had heard in meetings with studio executives that the Total Access promotion had hit Netflix hard in the fourth quarter. Antioco decided to go alone, to keep things informal and give Hastings privacy to say what was on his mind. He and Shepherd agreed to meet up for lunch as soon as Antioco wrapped things up with Hastings. Shepherd was disappointed to miss hearing what he hoped would be the admission that Netflix was in trouble because of the Total Access promotion.
As Antioco rode up into the hills above Park City’s crowded main streets to the address Hastings had indicated, he thought about the three years of gambles, sacrifices, and bravado it had taken to finally gain the upper hand against a foe whose success was riding on the demise of the video store. Still, it had cost him upwards of $500 million to rectify his mistake in not buying Netflix at $50 million when he had the chance. Netflix had gone on to spend a nearly equivalent amount to build and market its service in the ensuing years, but there was no disputing that its superior platform was the product of years of testing and tweaking and perfecting. The cab arrived, and Antioco paid and stepped out onto the porch of the steepled redbrick building where he had been let out. The Netflix chief answered the door himself. As soon as they sat down with their drinks, Hastings cut the pleasantries and turned serious.
He congratulated Antioco for the success of Total Access and admitted that the promotion had definitely gotten Netflix’s attention over the holiday quarter.
It was a great proposition—and Netflix couldn’t really match it, Hastings continued, but his analysts had calculated that each in-store trade was costing Blockbuster two bucks. With no limits on the number of movies Total Access subscribers—now three million strong and growing—could check out each month, Blockbuster’s debt would add up fast, Hastings observed.
Antioco waited for Hastings to come to the point.
The growth Total Access had created was phenomenal, but the only way to keep it up is by spending yourself into a corner, Hastings added. The minute you stop the free trades and raise your prices to make a profit, you lose your advantage, and we start growing again.
So what do you su
ggest? Antioco asked.
Let us buy your subscribers, Hastings said. We’re better at online rental—more technologically proficient.
Hastings made the proposal sound like he was doing Blockbuster a favor, but Antioco knew it was the closest he’d get to an admission that he had won.
I don’t know, Antioco said. I think we’re doing all right. Besides, I don’t know how we get a deal like that past the Federal Trade Commission’s antitrust department. Hastings tossed out the idea of forming a joint venture to buy the subscribers, to get around regulatory hurdles, and idly discussed other scenarios, but never mentioned money. They parted after agreeing to have their teams confer.
Antioco left the chalet with a sense of exultation.
• • •
SHEPHERD RETURNED FROM Sundance with his stomach in knots. His debrief with Antioco following the meeting with Hastings had evoked dueling emotions. The thrill of having Netflix admit that it could not compete with Total Access battled the dread of keeping the cash-burning promotion going indefinitely while Antioco and Evangelist figured out how to make it profitable.
The two months he had spent the previous year riding the company jet back and forth to Los Angeles with Zine to beg the studios to keep Blockbuster provisioned with newly released DVDs that it could not pay for had been extremely stressful. He did not want to repeat the experience.
He had to somehow force Antioco and Evangelist to confront the looming liquidity crisis posed by Blockbuster Online’s explosive growth.
With both companies pouring hundreds of millions of dollars into marketing online rental, the universe of online subscribers was expanding faster than anyone thought possible, to more than twelve million by the end of 2007. While Evangelist and Antioco crowed about having Netflix on the ropes, Shepherd wondered how Blockbuster would maintain quarterly earnings while handing out unlimited free rentals to two million new customers and suffering a painful contraction in store rental that was killing the Movie Gallery and Hollywood Video chains.