How Data Analysts Build Their First ML Model
For Data analysts moving into machine learning · Based on Codebasics Statistical ML Build Framework
// TL;DR
If you're a data analyst comfortable with spreadsheets and SQL but new to machine learning, the Codebasics Statistical ML Build Framework bridges the gap. It reuses skills you already have — auditing messy data, spotting inconsistent labels, understanding what a target column means — and adds the scikit-learn workflow: train_test_split, model.fit, model.predict, and evaluation with R², precision, recall, and confusion matrices. You'll learn to choose between regression and classification, avoid the accuracy trap on imbalanced data, and produce a trustworthy test-set benchmark. Use it to graduate from describing data to predicting outcomes with lightweight, interpretable models.
Why is this framework a natural next step for data analysts?
As a data analyst, you already do 80% of what machine learning demands: you audit messy data, reconcile inconsistent category labels like 'Male', 'M', and 'MALE', spot impossible outliers, and understand what each column means. The framework's 'Data Matters More Than the Model' principle confirms your instincts — in real projects, cleaning and preparing data is most of the work, and model training is a small fraction. The new skills are surprisingly few: defining a task type, splitting data, calling a couple of scikit-learn methods, and reading evaluation metrics.
How do I turn a report into a prediction problem?
Start by asking what you're trying to predict — this is your dependent variable (target). If it's a continuous number like revenue or delivery time, you have a regression problem. If it's a category like churn/no-churn or a product tier, you have classification. Everything else in your dataset that helps predict it becomes an independent variable (feature). In code, features go into X (a DataFrame, capital X) and the target goes into y (a Series, lowercase y). This single reframing turns your descriptive analysis into a predictive model.
How do I build the model step by step?
First, audit and clean the data the way you already know how — fix missing values, standardise inconsistent labels, convert numeric-looking strings, and remove impossible values. Then split the data with `train_test_split(X, y, test_size=0.2, random_state=42)`. The `random_state` guarantees reproducibility so your benchmarks are comparable across runs. Pick a simple algorithm to start: `LinearRegression` for continuous targets, `LogisticRegression` for categories. Train with `model.fit(X_train, y_train)`, then predict on unseen data with `model.predict(X_new)` — always passing a 2D array or DataFrame.
How do I know if my model is actually good?
Never evaluate on training data; that measures memorisation, not learning. Evaluate only on the held-out test set. For regression, use `model.score(X_test, y_test)` for the R² score, and compare train R² versus test R² — a large gap means overfitting. For classification, don't stop at accuracy. Print `classification_report(y_test, y_pred)` to see precision, recall, and F1 per class, and plot `confusion_matrix(y_test, y_pred)` to see exactly which errors the model makes. This is where your analyst eye shines: you can interpret which mistakes matter to the business.
What's the biggest trap to avoid?
The accuracy trap on imbalanced data. If you're predicting a rare event — churn, fraud, defects — a model that always predicts the majority class can score 99%+ accuracy while being useless. Your instinct to question a suspiciously good number is exactly right. Lean on recall for rare-event detection and inspect the confusion matrix before declaring the model ready.
What should I do next?
Pick one real dataset from your own work — a churn table, a sales log, a support ticket export — and run the full workflow end to end. Define the task type, clean the data, split it, train a simple model, and interpret the classification report or R² against a real business question. Don't just read this; write the code yourself. Only submit a model to production once you have a test-set benchmark you genuinely trust.
// FREQUENTLY ASKED QUESTIONS
Do I need advanced maths to use this framework as a data analyst?
No. You don't need to derive gradient descent or the sigmoid function by hand — scikit-learn handles the maths inside model.fit. You need to understand what the metrics mean: R² for regression, and precision, recall, F1, and the confusion matrix for classification. Your existing skills in data auditing and interpretation are more valuable than deep maths for getting started.
Can I use my existing SQL and spreadsheet skills here?
Absolutely. Your data-cleaning instincts transfer directly — spotting inconsistent labels, missing values, and impossible outliers is exactly the audit step the framework demands. You'll load data into a pandas DataFrame instead of a spreadsheet, but the mindset of scrutinising every column carries over. Since 80% of ML work is data preparation, your analyst background is a major head start.
Which algorithm should I start with for my first model?
Start simple: LinearRegression if your target is a continuous number, or LogisticRegression if it's a category. Establish a baseline test-set benchmark first. Only upgrade to Decision Trees, Random Forest, or XGBoost if the simple model's performance is insufficient. Starting simple keeps the model interpretable and gives you a clear yardstick to judge whether more complexity is worth it.