Going Viral Is Predictable With Machine Learning.

Using 36,000 posts to predict when YouTube videos will trend.

I was supposed to be in Barcelona.

I’d booked the week off work.

I needed the break.

I’d been working full-time during the week and completing my Master’s part-time during evenings and weekends for over a year.

Barcelona didn’t happen.

Instead, I spent nine days glued to my laptop.

Nine hours a day.

Writing code during the day.

Running the models overnight while I slept.

Waking up to check the results, and then back to the daily sprint.

But not because I had to.

Because I wanted to.

Why this machine learning project mattered.

This was the first real chance to test something I’ve been arguing for years:

Social media is not random - it’s predictable.

There are patterns inside the data.

Patterns that explain why some content succeeds, and the rest disappears.

The problem is that the tools most people use can’t see them.

Standard dashboards show you what happened.They don’t show you why it happened, or what should come next.

The experiment.

I got access to a dataset of over 36,000 real UK trending YouTube videos across 12 months.

Each video came with the usual data:

• When it was published

• Its title, description, and tags

• Views, likes, dislikes, and comments

• What category it belonged to

  
  

But with this data set, I could actually ask:

Can we use machine learning to predict how many days it will take for a video to trend?

Why raw data isn’t enough.

The numbers you see on a dashboard, likes, views, comments, are surface-level. 

So I built new measurements designed to capture the context behind the numbers:

• Like ratio (How many people liked the video compared to how many watched it.)

• Comment ratio (How many people commented compared to how many watched.)

• Channel history (How quickly that creator’s previous videos had trended.)

• Publish timing (How long after upload the video started gaining traction.)

• 
  

These aren’t complicated ideas.

They’re just one step further than most people look.

The result.

After a week of building, testing, and refining, the final model could predict when a YouTube video would trend within roughly five hours.

It explained over 94% of the variation in how long it takes for a video to trend.

That means almost all of the difference in the time it takes a video to trend could be explained by the patterns in the data.

Going viral is not random it is something that can be predicted.

Why this matters beyond YouTube.

Most social media teams still measure performance the same way:

• Open the dashboard.

• Check if the numbers went up.

• Hope next month looks better.

  
  

That’s not strategy.

But if we treat social media data with the same care as any business, financial, or behavioural dataset, by cleaning it properly, asking better questions, and looking for the patterns underneath the surface, we can build systems that actually learn and improve over time.

The bigger picture.

This project was one brick in a much larger wall.

It scratched the surface of what becomes possible when you stop treating each post as an isolated event and start seeing the patterns hiding within the data.

Those who can understand these patterns will change how effectively businesses influence and communicate with millions of people.

That’s when intelligently engineered content will be designed and optimised to rapidly change the world.

And also our very own individual lives.

- Mike