Antonio Sosa
  • Profile
  • Projects
  • Experience

Project 02 · Data Science

Transaction Intelligence & Customer Analytics

An MIA 5126 project for understanding historical banking behavior through relational data engineering, temporal analysis, customer segmentation, and responsible interpretation. The published system combines a reproducible Python workflow, read-only FastAPI service, and Angular dashboard.

Open live application ↗ View source ↗
1.05MTransactions across 72 months
4,500Account-level behavioral profiles
5Interpretable customer segments

Problem

Historical banking data is relational, time-dependent, and easy to misinterpret when analytical units or population coverage change. The goal was to create a trustworthy customer-intelligence foundation before attempting segmentation, forecasting, or supervised analysis.

The project uses the public Berka / PKDD’99 Financial Dataset and explicitly limits its outputs to historical behavior, portfolio-level activity, service co-occurrence, and analytical demonstrations—not fraud accusations or automated customer decisions.

System

From relational source to decision-ready insight

The pipeline validates eight relational source tables, reconstructs them in DuckDB, and produces Silver analytical tables and Gold presentation assets. The analysis covers monthly behavior, SARIMA forecasting, service associations, PCA, K-means and Gaussian mixtures, behavioral outliers, and a bounded loan-outcome case study.

A read-only FastAPI service exposes selected anonymous or aggregate outputs. An Angular dashboard presents five interactive views without exposing source account identifiers or arbitrary database access.

Evidence

The selected K-means solution produces five interpretable behavioral segments. The first two principal components explain 63.4% of variance, while the first four explain 86.2%. Random-start checks show stable cluster assignments when the documented skewness treatment is retained.

The SARIMA model improved holdout MAE relative to last-value and seasonal-naive baselines. Forecasts, association rules, outliers, and the small supervised case study are presented as method demonstrations with explicit limitations.

Notebooks

The complete four-part workflow is available below, from source-data validation through modeling and responsible interpretation.

Notebook 01Data Foundation & EngineeringProvenance, quality, relational joins, and architecture Notebook 02Behavioral & Temporal AnalysisFeatures, forecasting, and service associations Notebook 03Customer SegmentationPCA, clustering, stability, and outliers Notebook 04Validation & InsightsSupervised case study, error analysis, and limits

© 2026 Antonio Sosa

 
  • GitHub

  • LinkedIn