Data Science Degrees vs Portfolios: What Interviews Check

Two credentials, one gap

There is a recurring argument in data science circles about whether a degree or a portfolio matters more when applying for jobs. I think the argument is slightly misframed. A degree tells an interviewer that you sat through a curriculum and passed assessments under some form of supervision. A portfolio tells them you produced something and put it online. Neither of these facts is what gets tested in the room. What gets tested is whether you can reason about a problem you have not seen before, under mild pressure, in front of someone who already knows the right answer and is watching how you get there.

This distinction matters because candidates often over-invest in the wrong signal. Someone with a strong academic record may assume the transcript speaks for itself and turn up under-rehearsed on the practical side. Someone with a flashy portfolio may assume the repository speaks for itself and struggle when asked to explain a decision they made eight months ago. Both groups are optimising for the artefact rather than for the conversation that the artefact is meant to start.

Consider a concrete case. Two candidates apply for the same role. Candidate A has a master's degree with a dissertation on time series forecasting, scoring in the top decile of their cohort. Candidate B has no formal qualification beyond a short course, but has published three portfolio projects, including a churn prediction model with a clearly documented train and test split by customer signup date. In the interview, both are asked the same question: how would you validate a model that predicts which customers will cancel a subscription next month. Candidate A recites cross-validation textbook procedure without mentioning that customers who already cancelled should not leak into the training set via shared account features. Candidate B, drawing directly from their project, immediately raises the leakage issue because they hit it in practice and had to fix it. The degree did not fail candidate A, and the portfolio did not automatically save candidate B, but in that specific exchange, lived experience with a concrete pitfall won out over recited procedure.

What interviewers are actually listening for

In my experience preparing for and thinking about these conversations, three things tend to get checked regardless of which credential brought you to the table. The first is whether you understand your own evaluation. If you claim a model achieves a certain accuracy or F1 score, a good interviewer will ask how the split was constructed, whether there was any temporal or group leakage, and what the baseline was. A degree does not automatically answer this, because plenty of coursework uses pre-cleaned datasets with a single fixed split that nobody had to think hard about. A portfolio does not automatically answer this either, because plenty of portfolio projects use a random 80/20 split on data that should have been split by time or by entity to avoid the model implicitly memorising future information.

The second is whether you can justify a baseline. Interviewers are often less interested in your final complicated model and more interested in whether you tried something simple first and can say by how much your fancier approach beat it. If a candidate jumps straight to a gradient boosted ensemble without mentioning that a simple logistic regression or even a majority class predictor was tried and beaten by a small margin, that is a red flag, because it suggests the complexity was chosen for its own sake rather than justified by measured improvement. This is a place where portfolios have a structural advantage, because a public repository can show the baseline comparison explicitly in a way a transcript never will. But only if the candidate actually did the comparison and did not just present the final number.

The third thing checked is reproducibility of reasoning, not just of code. An interviewer might ask you to walk through what you would do differently if a key assumption in your project turned out to be false, for example if the data had a different missingness pattern in production than in the sample you trained on. This tests whether you understand why you made the choices you made, not just that you made them. Degree holders sometimes struggle here if their dissertation supervisor made key methodological choices for them. Portfolio builders sometimes struggle here if they copied a pipeline structure from a tutorial without internalising the reasoning behind each step.

Where each credential genuinely helps

None of this means degrees or portfolios are worthless signals; they are just indirect ones. A degree, when it involved a substantial independent project, can genuinely demonstrate sustained work on a single problem over months, including the messy middle period where nothing works and you have to diagnose why. That sustained diagnostic experience is hard to fake and hard to compress into a two-week portfolio sprint. It also usually exposes you to statistical foundations, such as why a p-value is not the probability the null hypothesis is true, which occasionally comes up directly in interviews for research-adjacent roles.

A portfolio, when it is built honestly rather than copied, demonstrates something a degree often cannot: end-to-end ownership. Choosing your own dataset, deciding your own evaluation protocol, hitting a data quality problem nobody warned you about, and writing up what you learned when the initial approach failed. Interviewers who ask you to talk through a project you built yourself are often listening for exactly this kind of ownership, because it correlates with how you will behave when a real production dataset does something unexpected six months into a job.

The honest answer, then, is that neither credential is checked directly. What gets checked is the quality of your reasoning about evaluation, your discipline about baselines, and your ability to explain why you did what you did rather than merely that you did it. A degree can give you exposure to the theory behind these habits. A portfolio can give you the scar tissue from actually practising them. The candidates who do best in interviews tend to be the ones who used whichever path they had access to as a forcing function to build those habits, rather than treating the credential itself as the finished product.

A practical takeaway

If you are choosing how to spend your preparation time before an interview, spend it rehearsing the story of one project you understand completely rather than polishing a list of qualifications. Be ready to state your baseline, your split strategy, and the one assumption that would break your result if it turned out false. That is the conversation that actually happens in the room, and it rewards depth on one honest example over breadth across a transcript or a repository list.