You Will Learn More From One Failed Kaggle Submission Than From Ten Completed Online Courses

Certificates feel like progress. Failure actually is.

Image by Gemini AI

There is a particular sense of accomplishment that one feels upon completing an online course.

The progress bar is at 100%, a certificate pops up, signed with your name, you screenshot it, you put it on LinkedIn, and you feel — honestly, truly — accomplished. Like you have progressed towards where you need to be. It isn’t my intention to deprive you of that feeling of accomplishment. What I wish to ask you, however, is this: after all of the courses and all of the certificates, can you actually do the job?

Not simply follow someone through what they’re doing, not pause every time they proceed too quickly. Open a blank notebook, be faced with a problem that is completely new for you, and find a solution.

The truth is, most people will not be able to answer with a resounding yes. This is not a matter of their intellect or willingness to learn. The problem lies in the very nature of data science education itself.

The Comfortable Lie of Structured Learning

Courses on Udemy and other MOOCs are great products. They are professionally made, logically structured, and taught by highly competent people. The most skilled teachers simplify the difficult and make it comprehensible.

That is precisely the problem.

If you were able to simplify things for yourself already, if the dataset was cleaned before, the problem already formulated and solved beforehand — you don’t learn data science; you learn how to perform it.

There is a name for this effect in psychology: illusion of explanatory depth. It has been proven time and time again that people think that they grasp things better once another person explains it to them. You watch the instructor importing data, dealing with missing values, creating the model, tuning the parameters, and it all seems so reasonable. It all works. You close the course, start working with your data, and no longer nod enthusiastically.

It’s not about the course. It is about the fact that understanding and ability are different things, and structured education can produce only the former.

What a Kaggle Competition Actually Does to You

For the first time, when you participate in a Kaggle competition for real, not just executing the starter notebook, but trying to make progress on the leaderboard, you will see a change.

No one tells you how to proceed. Data is unstructured in ways you could not learn from any modules you studied before. An evaluation metric is something you have come across but have not learned to optimize before. You find yourself competing with 4.000 teams, some of whom definitely know something you do not know.

You are forced to make certain decisions. Tough choices without an answer sheet available to verify whether you were right.

Should you normalize the variables or leave them alone? Can we apply a log transformation based on the data distribution pattern? Is our model over-fitting or is there variance in the scores of the leaderboard? How come my training metrics are looking great, but my score on the leaderboard isn’t? These are questions courses cannot teach you because courses, by nature, eliminate all such ambiguities. On Kaggle, you have more than enough of those.

The Specific Things Failure Teaches You

Let me give some concrete examples of what you learn from a failed submission, as “failure teaches you things” is vague advice at best.

What does an evaluation metric measure? If you’re optimizing for AUC-ROC in a particular competition and your score isn’t changing at all regardless of what you do, you suddenly start gaining a more profound understanding of what AUC-ROC actually measures and does not. And you won’t get it just from lectures.

Data cleaning is a complex process. Data preparation courses teach you that you need to fill gaps using mean, remove duplicates and encode categoricals. What they won’t teach you is how to handle missing values that weren’t randomly sampled; duplicates that look like near-duplicates; and categoricals with thousands of different values that weren’t mentioned in any tutorial you’ve ever read. And why would they?

How certain approaches will perform. You know that gradient boosting with default settings works better on tabular data than neural networks? Well, courses won’t necessarily tell you that. They will tell you how both of these approaches work; yet it will take practice to understand which of them is better in certain scenarios. You learn this through failure: try a certain algorithm, watch how it performs and read solutions of previous winners.

How to assess code quality of others. Kaggle discussions boards and submitted code are among the most valuable assets when learning about ML. However, you can benefit from them only after struggling with the problem. Before that, reading other people’s code is not as beneficial. It’s similar to reading the conclusion of a book before going through the whole story: interesting, but not as valuable.

Why We Keep Choosing Courses Anyway

Image by Gemini AI

If learning through competition is such a powerful method, why do people opt for taking courses instead?

It’s all about safety. Courses don’t evaluate. They never let you see the leaderboard and find yourself in the bottom 70 percent. They never force you to realize how much more sophisticated others’ solutions are than your own. You have a linear progression: minutes watched, percent finished, certificate achieved.

In contrast, Kaggle provides feedback that is brutally honest. You submit what you consider to be your best effort, and receive a score as an outcome. This score can be disappointing. At least it was, in my case, on the first few attempts.

People are hardwired to search for environments that make them feel competent, rather than environments where they’re actually competent. Courses create a fake sense of competence, and competitions generate genuine competence.

But humans often go for the former, since it provides more immediate gratification.

Another reason to opt for courses over competitions is a simple matter of completion bias. On Udemy, you have a progress bar and a clear milestone to achieve, which gives you a sense of accomplishment once reached. In Kaggle, there is no such thing.

The Right Way to Think About Courses

None of this is an indication that you should uninstall your Udemy app from your devices.

There are several reasons why people take courses; the reason here is that they provide you with a language and frame of reference. It is much faster to watch a well-explained video when you encounter concepts for the very first time, such as gradient boosting, regularization, or attention mechanisms.

The point is not that you shouldn’t be taking courses, but rather that you shouldn’t treat courses as practice for developing skills, which delays the actual process of working on the material.

Consider taking courses similarly to how a surgeon would consider anatomy books. It will not make you a surgeon to read through all the anatomy books; rather, you need knowledge about the body in order to understand what you are doing in the operating room.

Conclusion: If you have studied anatomy for several months but haven’t worked in an operating room yet, there is an issue with your study methods.

A Practical Path That Actually Works

And here is what such a learning path will look like:

Start doing competitions on Kaggle embarrassingly early. Do not hesitate to participate in a competition, even though you do not feel ready yet, since you will never feel fully ready. Choose one of the popular introductory competitions (Titanic, Spaceship Titanic, or House Prices, for example) and do them for a week prior to becoming qualified to do so. Confusion is the course.

Courses should be taken reactively rather than proactively. That means that when you run into something in a competition that you do not know about — a metric, some technique, or any other problem — that’s the time to look up the material on a relevant topic or read about that in one of those papers you did not read before. It works wonders to learn something immediately because you need it, compared to learning it just because it is the current lesson in your curriculum.

Obsessively study post-competition write-ups. Whenever any competition ends, the winners publish their approaches. These write-ups may seem to contain some of the most condensed, practical knowledge in machine learning, but there is one major caveat. The value of these write-ups is maximized if you completed in them yourself. Participating in any way changes the way you perceive the solution.

Write down what you did and why you did it, no matter what the result was. Logging actions during solving problems is crucial. Those who do not do that are likely to accumulate experience but not skills.

What “Failure” Actually Looks Like on the Other Side

Consider this: not many people who find themselves in the top 10% of a Kaggle challenge have done so in their very first attempts, nor their fifth. The leaderboard you’re looking at reflects many months’ worth of submissions that didn’t work out, fruitless trials at feature engineering, and models that looked promising but turned out not to be.

All those years later, what appears to be sheer talent is simply having the patience to endure the process.

It’s not the people who figured out how to circumvent the process altogether, but rather those who gave up trying to circumvent it. A submission that fails is not an obstacle to your journey of discovery. It is merely another piece of data. And when it comes to data, you can’t afford to settle for just one point.

The Certificate Won’t Save You

Image by Gemini AI

At some point in every data science interview, the certificates stop mattering.

Nobody’s going to ask you to walk them through your Udemy completion history. They’re going to give you a problem, or a dataset, or a case study, and they’re going to watch how you think.

That ability — to think through a messy, ambiguous, real problem without a tutorial guiding you — is built in the struggle. It’s built in the failed submissions, the debugging sessions that go nowhere, the models you had to throw away.

The certificate is a snapshot of what you consumed.

The competition is a record of what you survived.

Build the second kind of evidence. Start before you’re ready. Fail publicly, learn privately, and keep submitting.

The leaderboard will move eventually. More importantly, so will you.

If this hit something real for you, the best thing you can do isn’t share it — it’s open Kaggle and enter something today.