Why Data Science at all?
Data Science is not "AI magic", but rather careful questions to data:
What's going on? — descriptive analytics
Why is this happening? - diagnostic
What will happen next? - forecast
What should I do? - recommendations
Basic cycle: collect → clear → look (EDA) → test the hypothesis → model → explain the result and make a decision.
Rule #1: The result must lead to action. A number for the sake of a number is garbage.
Mini-cheat sheet: what to count and how to measure
Question | What we consider | Tool | Metric type | What it means |
|---|---|---|---|---|
How much do customers spend? | Average check (mean), median | Pandas/SQL | Descriptive | Basic revenue benchmark |
How "live" is the demand? | Conversion, frequency of purchases | Cohorts | Descriptive/diagnostic | Understanding the funnel |
Does the banner work? | Δ conversions, t-test/bootstrapping | SciPy/bootstrapping | Inf. statistics | Probability that the effect is not an accident |
What influences the response? | Correlations/feature importance | corr(), simple regression | Diagnostic | What factors are really significant |
How much will we sell tomorrow? | Forecast (linear/trees) | scikit-learn | Forecast | Optimize purchases/advertising |
Example 1. Coffee shop and average check
Scenario: the owner wants to understand why the revenue is "jumping".
Data: table orders (id, date, amount, payment method).
What we do:
We calculate average check and median.
We are building distribution of receipts.
We are looking day of the week — where are the peak days?
df['weekday'] = df['date'].dt.day_name()
avg_check = df['amount'].mean()
median_check = df['amount'].median()
by_weekday = df.groupby('weekday')['amount'].mean().sort_values(ascending=False)
Conclusions: if the average check >> the median, you have a long "tail" of expensive orders. Peaks on Fridays? So, it's better to launch the promo on Thursday night.
Example 2. Conversion from viewing to purchase
Scenario: online store: lots of views, few purchases.
Data: events view_product and purchase with user_id.
views = events.query('event=="view_product"').user_id.nunique()
carts = events.query('event=="add_to_cart"').user_id.nunique()
buys = events.query('event=="purchase"').user_id.nunique()
conv_view_to_cart = carts / views
conv_cart_to_buy = buys / carts
Conclusion: if the drawdown is at the "basket → purchase" stage, check the UX, form errors or shipping cost.
Example 3. Quick A/B test of a banner
from statsmodels.stats.proportion import proportions_ztest
count = np.array([buys_B, buys_A])
nobs = np.array([visits_B, visits_A])
stat, p = proportions_ztest(count, nobs)
Conclusion: even with a +0.8 p.p. difference, the confidence interval and the cost of traffic should be taken into account.
Example 4. Sales forecast "for tomorrow"
X = pd.get_dummies(df[['day_index','weekday']], drop_first=True)
y = df['sales']
model.fit(X_train, y_train)
pred = model.predict(X_test)
mae = np.mean(np.abs(pred - y_test))
Conclusion: forecast is needed to reduce uncertainty, not for "accurate guessing".
Example 5. What really affects the response?
Scenario: email newsletter. Why do some people open emails and others don't?
Idea: check correlations and logistic regression to assess the impact of features.
Data cleaning is half the battle
Problem | How to notice | What to do |
|---|---|---|
Passes |
| Delete or replace with median |
Emissions | Boxplot, z-score | Restrict by business rules |
Duplicates |
| Clear keys |
Different date formats | Parsing errors | Use |
Incomplete categories | Share < 1% | Merge into "Other" |
In Codice We provide short modules on Python and analytics with "real-life" tasks — CSV loading, cleaning, funnels, mini A/B test and the first model. No water and with clear results.
And we also have an active Telegram channel, where we discuss cool ideas, share experiences and analyze tasks together — learning becomes not only useful, but also fun.