Jay Frank: Algorithms Can’t Pick Hits

This post by Jay Frank, author of Futurehit.DNA and SVP of Strategy at CMT, is a response to Bruce Warila's recent post on the potential for computers to analyze hit songs.

image from farm5.static.flickr.com I’ve actively watched the discussion on algorithms being able to determine the hit worthiness of a song. I’ve debated the effectiveness of this with many of those involved with these services.

So I read with interest Bruce Warila’s recent take on these analysis. His conclusion was one that suggests there may be some merit in these services, but not to rely on them. In thinking through my arguments and my own attempts for predictive modeling, I actually have one big conclusion:

These Things Don’t Work.

In my attempts at finding future hits, I’ve found that the only thing you can accurately predict is whether a song would NOT be a hit. It’s pretty easy for anyone (even an algorithm) to spot a dud. But the variables for a hit are many. I learned this in great detail in 2004 while at Yahoo! with the artist Bonnie McKee (detailed in Chris Anderson’s book “The Long Tail”). At that point, every bell went off suggesting that her song “Somebody” would be a big teen hit. While posting huge results at Yahoo!, nobody else bit and the song failed. Interestingly, her talent became verified this year by having co-written three of this year’s biggest pop songs. But six years ago, it didn’t happen.

Why Algorithms Fail At Picking Hits:

  1. HITS STILL NEED GATEKEEPERS – Bonnie McKee didn’t happen because radio programmers, who would be needed to make the song a smash, didn’t “get it”. Never mind that all the data presented to them suggested their audience would respond favorably. Their opinion said otherwise and the gate was kept shut. If the same scenario happened today with the internet being more mature, Bonnie would have gotten more traction, but she would likely still have been hampered by the same issue. I see this happen with many songs on a weekly basis.
  2. HITS NEED THE RIGHT TEAM – The best song in the hands of the wrong team will fail to be a hit. Every time. Making a hit requires persistence, coordination, and momentum. Having the wrong people at the wrong time will kill a song. No algorithm can know if the head of promotion will have some hidden bias against a sound of a song. No algorithm can know if that head of promotion has lost respect by his team or by the radio programmers who he needs to convince a song is a hit. All things that can stop a song dead in its tracks.
  3. HITS NEED THE RIGHT TIMING – I’ve seen videos with snow in them released in May. I’ve seen a song have the bad fortune of being the sixth tempo song about partying by a solo male in the same month. I’ve seen labels squash a female singer’s single because the more popular female singer at the label also had a single with a similar vibe at the same time. All things that an algorithm can’t predict.
  4. HITS NEED THE RIGHT ATTITUDE FROM ALL WHO CREATED IT – I’ve seen amazing songs flop because the artist proceeds to act like a jerk at every promotional visit they attend. I’ve seen great songs tank because the manager pissed off someone at the label who instantly de-prioritized a song. I’ve seen surefire hits flop because a songwriter wouldn’t budge on splits. So many behind the scenes politics from the artist side can make a single cold in an instant. Occasionally, a song may be great enough to overcome these obstacles, but not often.
  5. LYRICS ARE TIMELY – I have not seen evidence that any of the algorithms analyze lyrics, but a well timed word placement is certainly a prime reason why many respond to a song. Then there’s the issue of slang. Would an algorithm pick up songs like “Bling Bling”, “Imma Be” or any song with the words “twerk”? Also, there’s the issue of clean vs. dirty versions and how that factors into making a song a hit in the past and its involvement in making a hit now since anyone can find dirty versions more easily. Does Cee-Lo’s “Fuck You” become a hit in the algorithm? Does the clean version “Forget You”? And for that matter, does the Glee version which recently debuted at #1 on the sales chart?
  6. AND, BY THE WAY, CAN SOMEONE DEFINE A HIT? – It’s nearly impossible to determine a hit. Most people (myself included) use the Billboard #1 as a benchmark. Yet many #1 hits are long forgotten today. Meanwhile, many other songs were never even singles to qualify, yet are remembered as hits. “Stairway To Heaven”? “How Soon Is Now”? Would you say Kyu Sakamoto’s “Sukiyaki” was a bigger hit than The Beatles’ “I Saw Her Standing There”? They were released only 1 year apart, yet by chart standards “Sukiyaki” is the bigger song. So who’s right?

And that’s the rub. Great strides in determining actual hits have been made in recent years, most notably by the Ultimate Chart. But the system in the past was extremely flawed. In order for algorithms to correctly gauge future hit worthiness based on past hits, one has to either accept the charts of the past (which were gamed to varying degrees) or bring subjectivity to the mix which defeats the purpose of the algorithm. Either one is not accurate enough for my tastes. Unless, of course, you’re just trying to predict that your song will NOT be a hit, but I think the blogosphere can do that for you for free.

Share on:


  1. Bruce writes about the song itself (as does “song science”), while Jay writes about the environment and circumstances surrounding the song. Apples and oranges.

  2. That’s pretty funny coming from a guy who has never picked a hit in his life. He worked for Yahoo correct? Not an A&R guy and far from a music person.
    At Yahoo he picked what he was told to pick.

  3. I never worked for Yahoo. I assume he’s talking about me or you Jay? If it’s me, the man should read my post prior to attempting to make a point.

  4. Also, “hit predicting” is the wrong use / wrong application for the technology.
    I will debate anyone (next week) that there are practical uses for the underlying technology.
    Also, from my post: “When it comes to songs, determining popularity potential (along a spectrum and within niches) and then matching songs to taste preferences, and artists to target audiences (through recommendation), are the technological advancements that should really matter to the majority of artists (IMHO).”

  5. I would presume he’s talking about me, which of course means he doesn’t know what he’s talking about. I was never told what to “pick”. We promoted based on audience response. And subsequently, created the base for many hits, which numerous labels would agree with.
    Bruce, I do agree with you that one might get some findings from the algorithms, but the only one I’ve seen effective is the “don’t waste your money” variety, personally. Now, granted, that has some merit, but would likely be ignored. And while we both agree that “hit predicting” is the wrong use, that’s how they’re largely marketed and that’s the net result most people expect so it’s only fair to judge it on those criteria.
    Great post, Bruce, and glad to see a debate created.

  6. The most significant line in this article is that “Hits still need gatekeepers”. This is both the a truth and a tragedy given how pathetic our AM/FM music radio programmers are today. Algorithms can select songs like that are similar to hits. They might be indicators of commercial viability (except for the lyrical aspect of the song) but cannot by any means predict if a song will be a hit or not.

Comments are closed.