Bootcamps vs Self-Taught Paths: What Portfolios Reveal

Why the credential is a distraction

Whenever I see the bootcamp versus self-taught debate, it tends to collapse into a proxy war about credibility. People want a shortcut for judging competence, and a twelve-week certificate or a GitHub streak both promise to be that shortcut. Neither is. I have looked at portfolios from both backgrounds and the split that actually predicts anything is not where someone learned, it is whether they understand what their own numbers mean.

A bootcamp graduate can produce a polished repository with clean commits, a tidy README, and a dashboard that looks production-ready. A self-taught learner can produce something scrappier but with a paragraph explaining exactly why they chose a time-based split instead of a random one. The second thing is worth more, even if it looks less impressive on first glance. Hiring managers who only skim for polish are optimising for the wrong signal.

This matters practically because bootcamps and self-study produce different failure modes, not different levels of ability. Bootcamps often optimise for shipping a finished-looking project within a fixed timeline, which rewards copying a known-good pipeline over understanding why each step exists. Self-teaching often optimises for breadth, which rewards collecting many small projects over stress-testing one properly. Both paths can produce someone excellent. Both paths can also produce someone who has never once asked whether their test set actually represents the future.

The tell: how leakage gets handled, or ignored

If I want a fast read on a portfolio, I look for one thing: does this person show awareness of data leakage, and did they do anything about it. Take a churn prediction project, a genre that appears in roughly every second portfolio. A common pattern is a random train-test split on customer-month rows, reporting an AUC of 0.94, and calling it done. That number is almost always too good, because rows from the same customer end up on both sides of the split, so the model partly memorises customer identity rather than learning the pattern of churn.

A portfolio that instead splits by customer ID, or better, by time, so that all of a customer's history sits entirely in train or entirely in test, and then reports an AUC of 0.79, is showing me something more valuable than the higher number ever could. It is showing me that the person understands the difference between a model that generalises and a model that has quietly cheated. The drop from 0.94 to 0.79 is not a failure, it is the moment the project became trustworthy.

The same pattern shows up with time series projects where someone forgets to lag their features properly and ends up using tomorrow's weather to predict today's ice cream sales, or with text classification where near-duplicate documents leak across the split. None of these mistakes are shameful on their own, everyone makes them early on. What matters for a portfolio is whether the writeup shows the mistake being caught and fixed, ideally with both numbers shown side by side. A single suspiciously perfect metric with no discussion of the split methodology is a bigger red flag than a mediocre metric with a clear explanation of how it was obtained.

This is where bootcamp curricula sometimes underserve people, because a tight schedule leaves little room to iterate on a project after the initial version works. Self-taught learners, paradoxically, sometimes have more room to revisit an old project a month later and fix the leakage they did not notice the first time. Neither path guarantees this happens, but a portfolio that shows revision history on evaluation choices is telling you something real about how the person will behave once the stakes are higher than a personal project.

Baselines and honesty do more work than model choice

The second signal I look for is whether a strong baseline appears before the fancy model. Suppose someone builds a gradient boosted model to predict house prices and reports a mean absolute error of eleven thousand pounds. On its own that number tells me nothing, because I do not know if eleven thousand is impressive or embarrassing. If the same person also reports that a simple median-price-per-postcode baseline achieves a mean absolute error of fourteen thousand, I suddenly know the boosted model is earning its complexity, even if modestly.

Without that baseline, I cannot tell the difference between a well-tuned model and an elaborate way of restating the average. This is not a minor stylistic preference, it is the difference between a project that demonstrates judgement and one that demonstrates familiarity with a library. Anyone can call a fit method. Fewer people stop to ask what the dumbest reasonable comparison would be and report it honestly, especially when the comparison makes their fancy model look only marginally better.

I have seen this baseline habit correlate more with intellectual honesty than with formal training. Some bootcamps teach it explicitly as part of the curriculum, which is genuinely useful, but plenty of self-taught learners pick it up from reading papers or from being burned once by an inflated result they could not defend in an interview. The habit is learnable regardless of path, and its presence in a portfolio is a much better predictor of how someone will behave with messy production data than any single algorithm they know how to implement.

Reproducibility rounds out the picture. A project with a pinned environment, a fixed random seed where it matters, and a note about which results are stable across seeds and which are not, tells me the author has thought about their own project as something other people might need to trust and rerun. This is a small thing to check for and it takes seconds, but it filters out a surprising amount of noise.

What this means in practice

If you are building a portfolio, spend less time chasing a tenth path and more time picking two or three projects and doing the unglamorous parts properly: a leakage-aware split, a stated baseline, and an honest account of what did not work. If you are evaluating someone else's portfolio, ignore the credential entirely and read for these signals instead. The route someone took to get here is a biography. The way they handle their own evaluation is a forecast, and it is the only part of the portfolio that reliably predicts what they will do with your data once you are not watching.