Exploratory correlation analysis identifying the behavioral factors most strongly associated with annual revenue per customer on a Latin American e-commerce platform.
The Growth & Retention team at NovaRetail+ needed to understand which customer behaviors drive annual revenue, in order to prioritize strategic initiatives for 2024 year-end planning.
Core business question:
What customer behavioral factors are most strongly associated with annual revenue generated per user?
Conduct a structured correlational analysis across 15,000 customer records to identify, quantify, and interpret the behavioral signals most relevant to revenue performance and to distinguish statistically significant associations from practically meaningful ones.
| Attribute | Detail |
|---|---|
| Records | 15,000 customers |
| Features | 12 variables |
| Missing values | None |
| Source | Internal platform data (NovaRetail+, 2024) |
Variable overview:
| Variable | Type | Description |
|---|---|---|
id_cliente |
Identifier | Unique customer ID |
edad |
Numeric | Customer age |
nivel_ingreso |
Numeric | Estimated customer income |
visitas_mes |
Numeric | Monthly platform visits |
compras_mes |
Numeric | Monthly purchases |
gasto_publicidad_dirigida |
Numeric | Targeted ad spend attributed to user |
satisfaccion |
Numeric | Satisfaction score (1–5 scale) |
miembro_premium |
Binary | Premium membership (1 = Yes, 0 = No) |
abandono |
Binary | Churn indicator (1 = Churned, 0 = Active) |
tipo_dispositivo |
Categorical | Device type (mobile, desktop, tablet) |
region |
Categorical | Geographic region (north, south, east, west) |
ingreso_anual |
Numeric | Target variable — Annual revenue per customer |
This project follows a structured EDA and correlation analysis pipeline:
1. Data Loading & Inspection Validated dataset structure, dtypes, null counts, and shape.
2. Data Cleaning
Corrected edad dtype from float64 to int64. No null imputation required.
3. Exploratory Analysis Descriptive statistics and distribution diagnostics across numeric, binary, and categorical variables.
4. Correlation Visualization Heatmap and scatterplot matrix to identify visual patterns before applying statistical methods.
5. Statistical Correlation Testing
| Method | Applied To |
|---|---|
| Pearson | Linear relationships between continuous variables |
| Spearman | Monotonic relationships with non-normal distributions or outliers |
| Point-Biserial | Binary vs. continuous variable pairs |
| Cramer's V | Categorical variable pairs |
6. Business Interpretation Five structured findings with visual evidence, numerical evidence, interpretation, and business implication.
7. Limitations & Next Steps Explicit acknowledgment of correlation vs. causation, large-sample p-value inflation, and Simpson's Paradox risk.
compras_mes is the only variable with a strong association with ingreso_anual (Pearson r = 0.967, p < 0.0001). Purchase frequency explains approximately 93.5% of the joint variation with annual revenue.
Business implication: Strategies targeting purchase frequency are the highest-leverage behavioral lever for revenue growth.
Targeted advertising spend shows a moderate monotonic association with monthly visits (Spearman r = 0.559, p < 0.0001). High individual-level variability limits its predictive strength.
Business implication: Advertising investment is positively linked to traffic, but returns vary significantly by customer segment.
miembro_premium shows no practically relevant linear association with any continuous variable. All Point-Biserial coefficients fall below |r| = 0.10.
Business implication: Premium status alone does not distinguish customers by behavior or revenue contribution in this dataset.
abandono shows negligible linear association with all continuous variables. The sole statistically significant pair (abandono vs satisfaccion, r = 0.024) has a negligible effect size.
Business implication: Linear correlation is insufficient to profile churned customers. Non-linear or clustering approaches are recommended.
miembro_premium vs abandono is the only categorical pair with a meaningful Cramer's V (V = 0.120, p < 0.0001). Low magnitude; not a robust churn predictor.
Business implication: Merits further investigation specifically a cross-tabulation of premium churn rates disaggregated by satisfaction level.
| Variable Pair | Method | Coefficient | Magnitude |
|---|---|---|---|
compras_mes vs ingreso_anual |
Pearson | r = 0.967 | Very Strong |
gasto_publicidad_dirigida vs visitas_mes |
Spearman | r = 0.559 | Strong |
compras_mes vs visitas_mes |
Pearson | r = 0.350 | Moderate |
ingreso_anual vs visitas_mes |
Pearson | r = 0.340 | Moderate |
ingreso_anual vs gasto_publicidad_dirigida |
Pearson | r = 0.200 | Weak |
miembro_premium vs abandono |
Cramer's V | V = 0.120 | Low |
| All other binary/categorical pairs | Point-Biserial / Cramér's V | < 0.10 | Negligible |
- Correlation ≠ Causation. No finding in this analysis supports a causal interpretation.
- Large sample effect. n = 15,000 produces significant p-values even for negligible effect sizes. Statistical significance and practical relevance are evaluated independently throughout.
- Simpson's Paradox risk. Global correlations have not been verified within subgroups (
region,tipo_dispositivo,miembro_premium). - Variables not included. Seasonality, promotional campaigns, and macroeconomic context are outside the scope of the available data.
Step 1 ” Simpson's Paradox verification
Disaggregate primary correlations by region, tipo_dispositivo, and miembro_premium. Prioritize the compras_mes vs ingreso_anual pair.
Step 2 ” Customer segmentation Apply K-Means or hierarchical clustering to behavioral variables to identify differentiated customer profiles. Evaluate whether global correlation patterns hold within segments.
Step 3 ” Predictive modeling
Develop a regression model using ingreso_anual as the dependent variable, with compras_mes, visitas_mes, and gasto_publicidad_dirigida as predictors. Evaluate multicollinearity before including correlated features simultaneously.
Technical
Python Pandas NumPy Seaborn Matplotlib SciPy
Exploratory Data Analysis Correlation Analysis Statistical Testing
Data Cleaning Data Visualization Analytical Documentation
Analytical
Statistical Thinking Effect Size vs. P-value Interpretation Business Insight Communication
Structured Problem Solving Findings Reporting Analytical Limitation Awareness
Analysis conducted for the Growth & Retention team, NovaRetail+ (2024)