Practical learning and project workspace

Free Roadmaps

Free Data Scientist Roadmap: Statistics, Machine Learning, and Deployment

A two-phase data science guide covering statistics, Python, pandas, SQL, machine learning validation, deep learning, deployment, portfolio projects, and career preparation.

DataUpto Editorial14 min read2,676 wordsResearch checked 27 October 2026

Free to learn. Build at your own pace.

The guide and linked videos are free to access. Software plans, certificates, advertising, API calls and cloud hosting may cost money. Use the official resource notes before signing up. Courses are curated for topic fit, not a verified ratings leaderboard.

1. What data science involves and how to use this roadmap

Data science combines statistical reasoning, programming, and domain knowledge to answer questions under uncertainty. Sometimes the right output is a model, but it may also be an experiment design, a careful descriptive analysis, or a recommendation not to automate. A strong practitioner asks how data was generated, whether the target can be measured reliably, and what happens when a prediction is wrong. Accuracy alone is not a complete definition of success.

This roadmap has two phases: foundations, then intermediate and advanced work. It moves through statistics, Python, pandas, SQL, machine learning, validation, introductory deep learning, and deployment. The sample schedule is a planning aid, not a promise of employment or mastery. Beginners may need substantially longer, particularly when learning mathematics and programming together. Advance when you can explain and reproduce the checkpoint, not merely when you have finished watching a lesson. Keep each project small enough that you can inspect its assumptions and errors yourself.

2. A free-first setup and realistic platform limitations

Install Python in an isolated environment and use a local editor or Jupyter notebooks. Learn how to create an environment, install packages, record versions, and restart a notebook before running it in order. NumPy, pandas, scikit-learn, and major deep-learning frameworks can be used without buying a course. SQLite is convenient for initial SQL practice; PostgreSQL adds a fuller database environment when you need it. Choose a simple stack rather than installing every available tool.

Hosted notebooks can help when local hardware is limited, but free tiers may have session limits, changing hardware availability, storage restrictions, or usage quotas. Google Colab's official FAQ describes such limitations; a free GPU is not guaranteed. Cloud hosting, managed databases, API calls, and persistent model serving can incur charges. Read current pricing, set spending controls where available, and do not confuse an educational credit with permanently free infrastructure. Never upload confidential employer or customer data to a hosted notebook without explicit authorization.

3. Phase one: Python skills that make experiments trustworthy

Learn numbers, strings, lists, tuples, dictionaries, sets, indexing, conditions, loops, functions, and modules. Practice comprehensions only after straightforward loops make sense. Understand mutable objects, default arguments, file handling, exceptions, and basic classes well enough to read library code examples. Learn to interpret tracebacks instead of repeatedly rerunning a failed cell. Use clear names and small functions so that data transformations can be reviewed and tested.

Write a script that loads a CSV, validates required fields, summarizes numeric columns, and saves a report. Add a command-line option for the input path and helpful errors for missing files. Practice Git locally to track source changes, but keep credentials and restricted datasets out of repositories. Your checkpoint is to run the same script on a second compatible file without editing its internals. Add tests for an empty input, a missing column, and an invalid numeric value. Reproducibility begins with reliable ordinary programming, not with the choice of model.

A collaborative learning workspace — Editorial illustration · Build skills through practice, not video completion alone.

4. Statistics and probability before sophisticated modeling

Learn distributions, mean and median, variance, quantiles, covariance, conditional probability, independence, and Bayes' rule. Use small examples to distinguish the probability of a positive test given a condition from the probability of the condition given a positive test. Study sampling, selection bias, measurement error, confidence intervals, hypothesis testing, effect sizes, and statistical power conceptually before relying on software output. A small p-value does not tell you that a result is practically important or free from bias.

Simulate repeated samples from a known distribution and examine how sample means vary. Compare a confidence interval with the observed distribution of individual values, because those answer different questions. Create an example where a misleading correlation appears after mixing groups. Your checkpoint is a notebook explaining a two-group comparison with an estimated effect, uncertainty, assumptions, and limitations. Include a note about multiple comparisons and why choosing only the most impressive result after many tests can overstate the evidence. Avoid causal language unless the design supports it.

5. Mathematics and NumPy for understanding models

Learn vectors, matrices, shapes, dot products, matrix multiplication, and the intuition of linear transformations. Understand a derivative as local change and a gradient as the direction of steepest increase. Learn how gradient descent updates parameters and why feature scale can affect optimization. You do not need to prove every theorem before modeling, but you should understand enough to notice incompatible array shapes and explain what a loss function measures.

Use NumPy to create arrays, inspect dtypes, index and slice values, broadcast operations, and calculate summary statistics. Implement a simple linear prediction and mean squared error with array operations. Then fit a small linear regression with a library and compare the predictions. Your checkpoint is to state the dimensions of the feature matrix, coefficient vector, and output before executing code. Explain why an enormous feature value can dominate a distance calculation and why preprocessing parameters must be estimated using training data rather than the entire dataset.

6. pandas, exploratory analysis, and data quality

Learn DataFrame indexing, data types, missing values, duplicates, groupby, aggregation, merges, reshaping, datetime operations, and string processing. Understand when a merge can multiply rows and how to validate expected key relationships. Keep raw data separate from cleaned outputs. Prefer transformations that can be rerun over manual spreadsheet edits whose history is difficult to reconstruct. Record units and category definitions in a data dictionary.

Explore a public dataset with a clear license: inspect schema, distributions, missingness by group, and potential measurement changes over time. Plot relationships without assuming that visible patterns will generalize. Create a reusable quality report and flag impossible values separately from merely unusual ones. Your checkpoint is to reproduce the cleaned dataset from raw inputs in a fresh session, with row-count checks at each major stage. Explain the consequences of your missing-value decisions, including whether missingness itself may carry information or reflect unequal data collection across groups.

7. SQL and the practical construction of modeling datasets

Learn filtering, aggregation, joins, CASE expressions, common table expressions, window functions, and date operations. A data scientist often spends significant effort defining the correct unit of analysis rather than tuning algorithms. Practice building one row per customer, transaction, device, or prediction event, depending on the task. Know the difference between an entity identifier and a meaningful predictive feature; an identifier may memorize records without helping on new cases.

Construct a modeling table from separate customers, events, and outcomes tables. Choose a prediction timestamp and allow only information available before that moment into the features. Calculate counts and averages within explicit historical windows. Your checkpoint is to explain the feature cutoff, outcome window, and row grain, then show that the joins preserve the intended number of examples. Test customers with no history and late-arriving records. A feature created from a future refund or a post-outcome support ticket can leak the answer even when the SQL is syntactically correct.

8. Phase two: classical machine learning and strong baselines

Begin with problem formulation: regression estimates a quantity, classification predicts a class or class probability, and clustering explores similarity without known labels. Learn linear and logistic regression, decision trees, random forests, and gradient-boosted trees before trying numerous unrelated algorithms. Understand regularization, underfitting, overfitting, and the bias-variance trade-off. For unsupervised work, learn how scaling and distance choices shape clusters and why a cluster is not automatically a meaningful business segment.

Use scikit-learn to build a dummy baseline, a simple linear model, and one tree-based model. Put preprocessing and estimation into a Pipeline, using a ColumnTransformer for different feature types. Compare candidates under the same validation design rather than giving the most complex model extra information. Your checkpoint is to beat or explain the baseline, inspect failure cases, and describe the cost of additional complexity. A simple model that is easy to maintain may be preferable to a marginally stronger one whose data dependencies cannot be supported.

9. Validation, leakage prevention, and metric selection

Separate training, validation, and final testing roles. Use random splits only when they reflect the intended deployment setting. Repeated observations from the same person or organization may require grouped splits, while future forecasting needs time-aware splits. Fit imputation, scaling, feature selection, and resampling inside the training process rather than before splitting. Tune hyperparameters using appropriate cross-validation and keep the final test set out of iterative decision-making.

Choose metrics around consequences. For regression, compare MAE and RMSE and inspect errors by target range. For classification, understand confusion matrices, precision, recall, ROC-AUC, precision-recall behavior, and probability calibration. Evaluate a threshold based on realistic costs or capacity rather than assuming 0.5 is always appropriate. Your checkpoint is a short validation memo explaining the split, baseline, metric, threshold, and leakage checks. Report subgroup and time-period performance where relevant, noting small samples and uncertainty. Do not describe one favorable test result as proof that a model will remain reliable forever.

A practical project workspace — Editorial illustration · Build skills through practice, not video completion alone.

10. Deep learning: learn the workflow without making it mandatory

Start deep learning after you understand training loops and validation with simpler models. Choose PyTorch or TensorFlow rather than learning both simultaneously. Study tensors, automatic differentiation, neural network layers, activation functions, loss functions, optimizers, batch size, epochs, and learning rate. Learn the roles of training and evaluation modes, regularization, and early stopping. Understand why validation data guides development but should not become an endlessly reused substitute for an untouched final test.

Build a small image or text classifier using a documented public dataset. Start with a compact network or transfer learning instead of training a large model from scratch. Track training and validation curves, compare against a simple baseline, and inspect confidently wrong predictions. Your checkpoint is to reproduce the experiment with saved configuration and weights, explain an overfitting pattern, and measure inference cost on available hardware. Deep learning is not required for every tabular problem, and buying expensive compute does not fix label leakage, poor sampling, or an undefined business objective.

11. Deployment, monitoring, and responsible operation

Deployment can begin as a batch script that reads new records and writes predictions; an API is not mandatory. Save the preprocessing pipeline together with the model and record its training data version, package environment, expected schema, and evaluation results. For an API exercise, use a lightweight framework, validate inputs, return understandable errors, and test both normal and malformed requests. Never load untrusted serialized model files, and keep secrets in an appropriate configuration mechanism rather than hard-coding them.

Measure latency and memory use, and plan what happens when a feature is unavailable. Track schema failures, missingness, prediction distributions, and performance when reliable labels arrive. Distribution change is a warning signal, not proof that model accuracy has fallen. Define rollback criteria and a human escalation path for consequential errors. Your checkpoint is a local demonstration with reproducible startup instructions, an example request, automated smoke tests, and a short monitoring plan. A public demo should use non-sensitive examples and conservative access settings, not expose an unrestricted expensive endpoint.

12. A flexible twenty-four-week study example

Weeks one through four cover Python, environments, Git basics, and a tested CSV analysis script. Weeks five through eight combine statistics, probability, NumPy, and simulation exercises. Weeks nine through twelve focus on pandas, SQL, exploratory analysis, and a reproducible modeling table. This completes the foundation phase only when you can explain data quality and sampling limitations, not just execute library calls.

Weeks thirteen through sixteen cover baselines, classical models, preprocessing pipelines, and validation. Weeks seventeen through twenty develop the main portfolio project, compare errors, and practice explaining results. Weeks twenty-one and twenty-two introduce a small deep-learning exercise. Weeks twenty-three and twenty-four package a model for local deployment, add monitoring notes, polish documentation, and rehearse interviews. Reserve weekly time for independent coding, mathematical review, and written explanation. If your schedule is crowded, extend the calendar rather than eliminating validation and testing. This is an illustrative sequence, not a deadline or a hiring guarantee.

13. Portfolio project: define a defensible predictive problem

Build a demand-prediction project using licensed public time-stamped sales data, or generate synthetic data and clearly label it. Define the prediction unit as a store-product-week and the objective as estimating next week's units using only information available at the forecast date. State who would use the prediction and what decision it informs. If stock availability is unknown, acknowledge that observed sales may understate demand during stockouts; the labels do not automatically represent unconstrained customer demand.

Step one is to record the source, license, schema, units, and time coverage. Step two is to profile missing weeks, duplicates, closures, and inconsistent identifiers. Step three is to create a complete calendar where appropriate and distinguish true zero sales from absent records. Step four is to design chronological training, validation, and test periods. Your checkpoint is a written problem statement and a diagram showing feature windows and prediction windows. Do this before training: a well-framed ordinary forecast is more convincing than an impressive metric on an ambiguous target.

14. Portfolio project: features, experiments, and delivery

Step five is to create lagged sales and rolling summaries that stop before the forecast period. Verify them manually for several rows. Step six is to compare a previous-period or seasonal naive baseline with a regularized regression and a tree-based model. Step seven is to evaluate overall error, error by store and product, and performance during unusual periods. Use MAE alongside a metric suitable for the decision, and be cautious with percentage errors when actual values approach zero.

Step eight is to inspect residuals and document where the model fails. Step nine is to package a batch prediction script with schema checks and a reproducible environment. Step ten is to publish a README, data-access instructions, feature definitions, validation diagram, experiment log, sample predictions, and a model limitations note. Include a small test verifying that changing future data cannot alter past features. Completion means another person can rebuild an allowed version of the experiment, understand why the baseline matters, and see how the model would be monitored after delivery.

15. Pakistan and global opportunities: show evidence, not inflated titles

For opportunities in Pakistan, examine data scientist, junior machine learning, risk analytics, forecasting, research assistant, and analytics roles individually. Some positions emphasize reporting or SQL while others expect stronger software engineering or postgraduate research experience. Tailor your application to the actual responsibilities and honestly describe your level. A relevant forecasting project with clear validation can be more persuasive than listing every framework you have briefly imported.

For global remote work, check location eligibility, time zones, contracting terms, and requirements for handling customer data across borders. Show written communication through a concise project report and a reproducible repository. Consider analyst or data engineering adjacent work when it matches your current skills; there is no requirement to obtain a particular title immediately. For freelancing, agree on the decision problem, available labels, baseline, acceptance metric, deployment scope, and maintenance responsibilities. Do not guarantee a target accuracy before examining lawful, representative data, and do not reuse private client datasets in a public portfolio without explicit permission.

16. Interview FAQ and the final readiness checkpoint

What mathematics should you explain? Be comfortable with probability, uncertainty, linear algebra shapes, derivatives, and what optimization minimizes. How do you handle a model that performs well offline but poorly in production? Investigate leakage, train-serving mismatch, population changes, label delays, and inappropriate validation before changing algorithms. How do you discuss an unsuccessful experiment? State the hypothesis, design, evidence, and what you learned rather than concealing the result. Practice SQL and Python exercises while explaining correctness and edge cases.

Do you need deep learning for every data science job? No; requirements depend on the problem and employer. Does a certificate guarantee readiness? No; independently reproducible work and sound reasoning matter. When should you apply? When you can demonstrate the relevant requirements honestly, even while continuing to learn. Your final checkpoint is to explain a project end to end: data provenance, target definition, split design, baseline, metrics, uncertainty, errors, deployment, and limitations. If any link is missing, improve that part before adding another algorithm. Responsible judgment is a central skill, not an optional finishing touch.

Luke Barousse

Dashboards, cleaning and modeling. Video free; supporter practice/community/certificate optional paid. Official course: https://www.lukebarousse.com/powerbi

Watch on YouTube

Free resources & official downloads

Tableau sample datasets

Free practice data; check individual dataset terms.

Python downloads

Free open-source interpreter; choose a supported release compatible with libraries.

PostgreSQL downloads

Free local database; MySQL-based course syntax may need adaptation.

pandas documentation

Free getting-started guides and API reference; older video APIs may differ.

scikit-learn documentation

Free user guides, examples and installation links.

Explore free datasets

Sources & further reading

Career examples are guidance, not vacancy announcements or job guarantees. Confirm requirements, regional access and current product terms at the linked official source.

Choose your next practical step

Ask Qasim Anwar for guidance on your learning path, projects or career applications.

A detailed two-phase digital marketing roadmap covering research, copy, SEO, analytics, Google and Meta campaign planning, consent-aware email, CRM, experimentation, and an honest no-spend portfolio.

15 min read10 video resourcesRead guide