Start with Problems, Not Tools

Understand why data science exists before collecting libraries and certificates.

Learn the difference between descriptive, diagnostic, predictive, and prescriptive questions. Practise converting broad requests such as “improve customer retention” into measurable questions: Which customers are leaving? When does the risk increase? Which factors are associated with churn? What intervention can be tested?

Learn Python and Reproducible Workflows

Use programming to inspect, transform, analyse, and automate data work.

Learn variables, data types, conditions, loops, functions, collections, modules, exceptions, and file handling. Then move to NumPy for numerical arrays, pandas for tabular data, Matplotlib and Seaborn for visualisation, and Jupyter notebooks for exploratory work. Learn Git fundamentals so that changes to code and analysis can be tracked.

import pandas as pd

sales = pd.read_csv("sales.csv")
sales["revenue"] = sales["quantity"] * sales["unit_price"]
summary = sales.groupby("region", as_index=False)["revenue"].sum()
print(summary.sort_values("revenue", ascending=False))

Build Mathematics and Statistics Intuition

Learn enough mathematics to understand data behaviour and model decisions.

Prioritise descriptive statistics, probability, distributions, sampling, correlation, linear algebra intuition, and optimisation concepts. You should understand why a formula is used, what assumptions it makes, and how to interpret the result. Memorising symbols without applying them is not the goal.

The arithmetic mean is:

Variance measures dispersion around the mean:

For model evaluation, mean squared error is commonly expressed as:

Master SQL and Relational Data Thinking

Most organisational data lives in systems that must be queried and joined correctly.

Learn SELECT, WHERE, CASE, GROUP BY, HAVING, joins, subqueries, common table expressions, window functions, date operations, and null handling. Also understand tables, keys, relationships, granularity, duplicate rows, and why an incorrect join can silently produce a misleading result.

SELECT
    c.customer_segment,
    COUNT(DISTINCT o.customer_id) AS active_customers,
    SUM(o.order_amount) AS total_revenue
FROM customers AS c
JOIN orders AS o
    ON c.customer_id = o.customer_id
GROUP BY c.customer_segment
ORDER BY total_revenue DESC;

Learn Data Cleaning, EDA, and Visualisation

Before modelling, understand whether the data is trustworthy and what it represents.

Study missing values, duplicates, invalid categories, inconsistent units, date parsing, outliers, leakage, sampling bias, and data-quality rules. During exploratory data analysis, examine distributions, relationships, segments, trends, anomalies, and possible explanations. Visualisation should answer a question—not merely decorate a report.

Develop Statistical Reasoning and Experimentation Skills

Use samples carefully and avoid claiming more than the evidence supports.

Learn sampling, estimators, confidence intervals, hypothesis testing, effect size, statistical power, multiple comparisons, regression interpretation, and experiment design. A p-value does not measure business value, and statistical significance does not automatically establish practical importance or causality.

Learn Machine Learning in the Correct Order

Build baselines first, then use complexity only when evidence justifies it.

Begin with linear regression, logistic regression, decision trees, ensemble methods, nearest neighbours, and clustering. Study preprocessing, feature engineering, train-validation-test separation, cross-validation, hyperparameter selection, class imbalance, overfitting, underfitting, and leakage.

from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_absolute_error
from sklearn.ensemble import RandomForestRegressor

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)
model = RandomForestRegressor(random_state=42)
model.fit(X_train, y_train)
predictions = model.predict(X_test)
print("MAE:", mean_absolute_error(y_test, predictions))

Build End-to-End and Production-Aware Projects

A notebook is an experiment; a dependable solution needs an operational lifecycle.

Learn modular code, configuration, testing, packaging, data and model versioning, APIs, containers, batch versus real-time inference, monitoring, drift, rollback, security, and documentation. You do not need to become a full MLOps engineer immediately, but you should understand what happens after a model is approved.

Choose a Specialisation Only After the Core

Specialisation is most valuable when built on strong foundations.

• Natural language processing: text classification, embeddings, retrieval, evaluation, and language-model applications
• Time-series analysis: trend, seasonality, forecasting, backtesting, and temporal leakage
• Computer vision: image classification, detection, augmentation, and transfer learning
• Causal inference: experiments, observational studies, confounding, and treatment effects
• Recommendation systems: collaborative filtering, ranking, feedback loops, and evaluation
• Generative AI: prompting, retrieval-augmented generation, agents, evaluation, safety, and cost
• Analytics engineering: reliable transformation models, testing, lineage, metric definitions, and warehouses

Build a Portfolio That Demonstrates Judgment

A strong portfolio explains decisions, not only code.

Build a small number of complete projects rather than many copied notebooks. Each case study should describe the problem, stakeholder, data source, data quality, assumptions, method, baseline, evaluation, findings, limitations, ethical considerations, deployment idea, and business recommendation.