Free Data Scientist Roadmap: Statistics, Machine Learning, and Responsible Deployment Quantum Institute of Analytics (QIA) — original free learning checklist Reviewed 27 October 2026 Learning time is flexible. Software, API use and examinations may cost extra. [ ] 1. Start with a prediction problem worth solving Data science combines statistical reasoning, software, and domain knowledge to investigate patterns and build systems that support decisions. It is not simply fitting a complex model to a downloaded table. Before predicting anything, identify who would act on the output, when they would act, and what errors would cost. For example, forecasting tomorrow's demand for a fictional pharmacy has different requirements from explaining last year's sales. One needs future-valid evaluation; the other may need careful descriptive analysis rather than machine learning. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 2. Set up a free learning environment and sustainable routine Python, Jupyter, pandas, scikit-learn, and many scientific libraries are open-source tools that can run locally without software subscription fees. That does not make every surrounding service free. Hosted notebooks can impose session limits, data restrictions, paid hardware options, or changing account requirements. Cloud deployment can incur charges even when learning documentation is freely accessible. Check the terms of each service, avoid entering payment details casually, and begin with datasets small enough for a regular laptop. A graphics processor is unnecessary for the classical machine-learning curriculum here. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 3. Learn Python foundations and basic software habits Study variables, numeric types, strings, lists, dictionaries, control flow, functions, modules, exceptions, and file handling. Practice writing a function that validates a transaction amount and another that summarizes a list of delivery durations. Understand mutability and missing values before manipulating large tables. Learn to read tracebacks from the bottom upward, reduce a failure to a small example, and consult the relevant official documentation. Memorizing syntax is less useful than being able to explain what a function accepts, returns, and assumes. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 4. Build statistical intuition with simulation Study populations, samples, random variables, distributions, expectation, variance, covariance, and conditional probability. Connect each idea to an actual question. A sample of completed deliveries excludes canceled orders, which may be the very cases a manager cares about. Averages hide variation: two routes can have the same mean travel time but very different probabilities of extreme delay. Use simple simulations to see how sample means vary across repeated samples rather than treating every reported average as a fixed truth. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 5. Learn the mathematics that clarifies models Develop practical linear algebra: vectors, matrices, dot products, matrix multiplication, norms, and projections. Connect a feature matrix to rows representing observations and columns representing measured attributes. Understand why a coefficient vector must have a compatible shape. Learn derivatives as local rates of change, gradients as directions across several parameters, and optimization as repeated improvement of an objective. You do not need to finish an entire mathematics degree before using a regression model, but you should recognize what the algorithm is minimizing. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 6. Use pandas and SQL to create trustworthy datasets Learn pandas indexing, type conversion, missing-value handling, groupby operations, joins, reshaping, and date processing. Add SQL selection, aggregation, joins, and window functions so that you can retrieve training data deliberately. Inspect row counts and key uniqueness after every merge. When customer attributes change over time, joining today's customer profile to last year's orders can accidentally introduce future information. Record what each feature means and the earliest moment at which it becomes available. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 7. Explore data without contaminating your evaluation Exploratory analysis should reveal distributions, missingness, unusual values, and plausible relationships. Plot target frequency over time and compare relevant subgroups, but separate exploration from final evaluation. Reserve an appropriate holdout before using outcomes to make repeated modeling decisions. If you inspect final test results and then choose features to improve those results, the test set has become part of training in a broader sense. Your reported performance will be more optimistic than an untouched evaluation would justify. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 8. Build baselines and understand core model families Begin with a dummy classifier or regressor and a simple operational rule. Then fit linear or logistic regression and a small decision tree. Learn how regression predicts continuous quantities and classification estimates categories or probabilities. Study overfitting by comparing training and validation performance as model complexity changes. Add random forests or gradient boosting only after you can explain the baseline, target, and evaluation design. More complex models can improve performance, but they also create new tuning, interpretation, and maintenance burdens. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 9. Match validation to the way predictions will be used Random splitting is not universally valid. For future demand forecasting, use chronological evaluation and avoid training on records that occur after the evaluation period. For repeated customers, patients, stores, or machines, consider grouped splitting when the deployment goal involves unseen groups. If the use case involves both future time and new locations, design an evaluation that reflects both constraints as closely as the available data permits. Explain remaining gaps instead of claiming one split proves universal generalization. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 10. Choose metrics, thresholds, and uncertainty honestly For regression, compare mean absolute error, root mean squared error, and a useful baseline; inspect error by target range and subgroup. Percentage errors can behave badly near zero. For classification, learn confusion matrices, precision, recall, and precision-recall tradeoffs when positive events are rare. Accuracy can look excellent when a model simply predicts the majority class. Distinguish ranking quality from calibrated probabilities, and check whether predicted probabilities roughly match observed event frequencies. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 11. Learn interpretation, fairness, and causal boundaries Inspect coefficients when their interpretation is justified, use permutation importance carefully, and explore partial-dependence style summaries with awareness of correlated features. Importance is not causation. A feature may predict delay because it reflects an existing scheduling policy, not because changing that feature would improve outcomes. Explanations can also become unstable when multiple correlated inputs contain similar information. Describe the technique and its assumptions rather than labeling any plot a complete explanation of model behavior. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 12. Deploy a small model with realistic safeguards Deployment means making a model usable through a repeatable interface, not merely saving a notebook. Build a small local API using a documented Python framework, validate incoming fields, and return predictions with a model version. Save the complete preprocessing pipeline with the estimator. Test malformed requests, unseen categories, missing fields, and out-of-range values. Do not load serialized model files from untrusted sources, since some serialization formats can execute code when loaded. Document the intended Python and library versions. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 13. Create a portfolio with two distinct scientific questions Build one supervised learning case and one statistical or forecasting case. For the first, use synthetic delivery risk data with realistic feature availability and a transparent generator. Make clear that synthetic performance does not establish real-world usefulness. For the second, use appropriately licensed public time-series data or an experiment simulation. Compare simple seasonal or historical baselines before adding sophistication. Each project should include a problem brief, source record, reproducible preparation, evaluation design, results, limitations, and a short audience-specific recommendation. Checkpoint: build a small example, explain its limitations, and record evidence. [ ] 14. Use self-assessment to choose the next learning cycle Common mistakes include learning deep learning before understanding validation, tuning on the test set, treating a correlation heatmap as an explanation, and publishing notebooks that cannot run from a clean session. Another frequent error is assuming more data automatically solves biased sampling or inconsistent labels. Keep an experiment log with the hypothesis, change, result, and decision. When an experiment fails, preserve the lesson rather than deleting every unsuccessful result from the narrative. Checkpoint: build a small example, explain its limitations, and record evidence. FREE RESOURCES Python foundations: https://docs.python.org/3/tutorial/ NumPy foundations: https://numpy.org/doc/stable/user/absolute_beginners.html SciPy statistical learning: https://docs.scipy.org/doc/scipy/tutorial/stats.html scikit-learn model evaluation: https://scikit-learn.org/stable/modules/model_evaluation.html scikit-learn validation: https://scikit-learn.org/stable/modules/cross_validation.html FastAPI local deployment: https://fastapi.tiangolo.com/tutorial/ Kaggle Learn: https://www.kaggle.com/learn UCI Machine Learning Repository: https://archive.ics.uci.edu/ Google Machine Learning Crash Course: https://developers.google.com/machine-learning/crash-course Python Data Science Handbook: https://jakevdp.github.io/PythonDataScienceHandbook/ DIRECT VIDEO LESSONS The ONLY Data Science Roadmap You Need to Get a Job: https://www.youtube.com/watch?v=_4RVudwQL38 Python For Data Science Full Course | Data Science With Python Full Course In 12 Hours | Simplilearn: https://www.youtube.com/watch?v=nHAPcZRg9VM Learn Data Science Tutorial - Full Course for Beginners: https://www.youtube.com/watch?v=ua-CiDNNj30 AI Foundations Course – Python, Machine Learning, Deep Learning, Data Science: https://www.youtube.com/watch?v=0oyDqO8PjIg Learn Python - Full Course for Beginners [Tutorial]: https://www.youtube.com/watch?v=rfscVS0vtbw StatQuest: Logistic Regression: https://www.youtube.com/watch?v=yIYKR4sgzI8 Linear Regression, Clearly Explained!!!: https://www.youtube.com/watch?v=7ArmBVF2dCs Complete Python Pandas Data Science Tutorial! (Reading CSV/Excel files, Sorting, Filtering, Groupby): https://www.youtube.com/watch?v=vmEHCJofslg Machine Learning with Python and Scikit-Learn – Full Course: https://www.youtube.com/watch?v=hDKCxebp88A Scikit-Learn Full Crash Course - Python Machine Learning: https://www.youtube.com/watch?v=SIEaLBXr0rk Building makemore Part 4: Becoming a Backprop Ninja: https://www.youtube.com/watch?v=q8SA3rM6ckI Building makemore Part 2: MLP: https://www.youtube.com/watch?v=TCH_1BHY58I