#Bank Marketing Insights — Data Cleaning & Correlation Analysis

A bank's marketing team needed to know which clients were actually worth calling. This project cleans and analyzes their campaign contact data — client demographics, financial indicators, campaign mechanics, and economic context — to surface which factors correlate with a successful outcome, and to be honest about which ones don't hold up statistically.

Takes raw bank marketing campaign data through a full cleaning and encoding pipeline, then runs a correlation analysis against campaign outcome to identify the strongest signals. Along the way, it tests whether the cleaned data would actually support a Multiple Linear Regression model — and when the linearity assumption fails, it stops there rather than reporting a model built on a broken assumption.

• Normalized inconsistent categorical formatting (job and education fields used periods instead of underscores)
• Converted unknown sentinel values in credit_default and mortgage to proper NaN, then quantified missingness (~20% missing in credit_default) instead of silently imputing
• Reconstructed a proper last_contact_date from separate month/day fields, with month cast to an ordered categorical
• Label-encoded previous_outcome; one-hot encoded job, marital, and education for correlation analysis
• Split the cleaned dataframe into client, campaign, and economic-indicator subsets scoped to their respective analysis questions

A Pearson correlation matrix against the binary campaign_outcome target surfaced the strongest signals:

With features encoded, the next step was testing whether the data actually supported Multiple Linear Regression. It didn't — the linearity assumption between predictors and the target failed diagnostic checks. Rather than report a regression built on a broken assumption, the analysis stopped at correlation-based insight extraction and flagged predictive modeling as future work requiring techniques that don't assume linearity.

• Contact duration is the single strongest signal — longer calls track strongly with success, suggesting reps should be coached toward sustained engagement over rushing through calls.
• Repeat engagement compounds — clients with more previous campaign contacts and a positive previous outcome convert at meaningfully higher rates, making re-engagement campaigns a high-leverage target.
• Segment-level patterns matter — students and retirees showed the highest average contact duration; tailoring messaging by occupation, age, and marital status (e.g., stability-focused messaging for older clients, flexibility-focused for singles) is a natural next step.

Pipeline (single-notebook, no live service):

Raw campaign CSV (client + campaign + economic attributes)
  → Categorical normalization (job/education formatting fixes)
  → Missing-value handling (unknown → NaN, quantified: ~20% missing in credit_default)
  → Date reconstruction (month/day → ordered categorical → last_contact_date)
  → Label encoding (previous_outcome) + one-hot encoding (job, marital, education)
  → Correlation matrix against campaign_outcome
  → Linear regression assumption check (linearity) → failed → regression abandoned
  → Insight extraction + segment-level analysis (role, age, education, marital status)

Key engineering decisions:

• Stopped short of forcing a linear model — after encoding features for Multiple Linear Regression, the notebook explicitly tested the linearity assumption between predictors and the target and found it didn't hold. Rather than ship a model that violated its own statistical assumptions, the analysis was scoped down to correlation-based insight extraction — a deliberate call that a plausible-looking wrong model is worse than an honest correlation study.
• Quantified missingness before deciding how to handle it — unknown values in credit_default/mortgage were converted to NaN and measured (~20% missing in credit_default) rather than silently imputed, so downstream analysis could account for the gap.
• Split into client/campaign/economic subsets — separating the encoded dataframe by domain keeps each analysis question scoped to the relevant columns instead of one wide, unfocused table.

Next architectural step: since linearity assumptions failed, tree-based models (Random Forest, Gradient Boosting) that don't require linear relationships are the natural next step for actual campaign-outcome prediction — not attempted here.