Data Science, Step by Step

1. Frame the problem

Every project starts with a clear question. Translate a business need ("why are customers leaving?") into a measurable target ("predict which customers will cancel in the next 30 days").

Decide up front how success will be measured — accuracy, revenue impact, time saved — so you know when the work is done.

Who will use the result, and how?
What decision will change because of it?
What data exists, and what would you need to collect?

2. Collect and clean the data

Real-world data is messy. Expect missing values, duplicates, inconsistent formats and outliers. Cleaning usually takes more time than modelling — and matters more.

Missing values: drop, fill with a statistic (median/mode), or model them
Duplicates: identify by a key and remove
Types: parse dates, convert numeric strings, standardise categories
Outliers: investigate before removing — they may be the signal

python

import pandas as pd

df = pd.read_csv("customers.csv")
df = df.drop_duplicates(subset="customer_id")
df["signup_date"] = pd.to_datetime(df["signup_date"])
df["monthly_spend"] = df["monthly_spend"].fillna(df["monthly_spend"].median())
df.info()

3. Explore the data (EDA)

Exploratory data analysis builds intuition. Look at distributions, relationships between variables and differences between groups before you build anything.

Summary statistics: mean, median, standard deviation, percentiles
Distributions: histograms and box plots
Relationships: scatter plots and correlation matrices
Segments: group-by comparisons (e.g. churn rate by plan type)

4. Statistics you will use constantly

Mean vs median — the median is robust to outliers
Variance and standard deviation — how spread out values are
Correlation — strength of a linear relationship (not causation)
Sampling and the central limit theorem — why averages of samples behave predictably
Hypothesis testing and p-values — is a difference likely to be real or chance?
Confidence intervals — a range of plausible values for an estimate

5. Build a model

Choose the simplest model that answers the question. Start with a baseline (e.g. always predicting the most common class) so you can tell whether a model actually adds value.

Regression — predicting a number (price, demand)
Classification — predicting a category (churn / no churn)
Clustering — finding natural groups without labels

python

from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression

X = df[["tenure_months", "monthly_spend", "support_tickets"]]
y = df["churned"]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

model = LogisticRegression(max_iter=1000).fit(X_train, y_train)

6. Evaluate honestly

Always evaluate on data the model has never seen. Pick metrics that match the problem — accuracy can be misleading when one class is rare.

Regression: MAE, RMSE, R²
Classification: precision, recall, F1-score, ROC-AUC
Use cross-validation for a more stable estimate
Watch for overfitting: great training scores, poor test scores

7. Communicate the result

A model nobody understands rarely gets used. Lead with the answer, show one or two clear charts, explain limitations and recommend a next action.