Customer Segmentation, Purchase Behavior & Store Trial Impact Evaluation
This repository presents a two-part retail analytics case study designed to replicate the type of analytical work performed within commercial analytics and strategy teams. The project focuses on understanding customer-driven demand patterns and rigorously evaluating the effectiveness of an in-store layout intervention using causal inference principles.
Rather than treating analysis as an end in itself, this project emphasizes decision-oriented analytics—linking data exploration, statistical validation, and business interpretation to concrete commercial recommendations.
The work is intentionally structured to reflect how analytics outputs are consumed in practice:
- Exploratory analysis to understand what is happening
- Structured validation to determine why it is happening
- Synthesized reporting to support what should be done next
├── chips_category_analysis.ipynb
│ └── Customer segmentation, purchase behavior analysis, and demand driver identification
│
├── store_trial_evaluation.ipynb
│ └── Trial vs control store selection, pre-trend validation, uplift estimation, and uncertainty analysis
│
├── quantium_analysis.pptx
│ └── Executive-level report structured using the Pyramid Principle
│
└── README.md
The notebooks contain the full analytical workflow, while the presentation distills findings into a format suitable for senior business stakeholders.
Retail organizations regularly face two recurring analytical challenges:
Sales growth can emerge from multiple underlying mechanisms:
- Acquiring more customers
- Increasing how often existing customers purchase
- Increasing spend per transaction
Without separating these effects, organizations risk making incorrect decisions—for example, investing heavily in price promotions when demand is actually driven by habitual purchasing behavior.
This project addresses the question:
What truly drives chip category sales across different customer life stages?
Store layout changes are costly to implement and difficult to reverse. As a result, decision-makers need high confidence that any observed performance improvement is:
- Causally linked to the intervention, and
- Not driven by seasonal variation, noise, or unrelated demand shifts
This project addresses the question:
Did the layout change create incremental value beyond what would have occurred naturally?
The analysis follows three guiding principles:
-
Decomposition over aggregation
Sales are decomposed into customer count, transaction frequency, and spend behavior rather than analyzed as a single metric. -
Causal defensibility over correlation
Store performance is evaluated using trial–control logic and pre-trend validation rather than naive before–after comparisons. -
Decision relevance over metric completeness
Only analyses that materially influence a business decision are emphasized in documentation and reporting.
📓 chips_category_analysis.ipynb
The first phase of the analysis focuses on understanding who drives chip sales and how. Rather than simply ranking segments by revenue, the analysis investigates why certain customer groups contribute more to total sales.
- Are higher sales driven by a larger customer base or higher purchase frequency?
- Which customer life stages demonstrate consistently repeat-driven behavior?
- Which segments offer the most stable and predictable demand?
Answering these questions is essential for informing assortment planning, promotional strategy, and store layout prioritization.
This visualization compares customer segments along two critical dimensions:
- Purchase frequency, which reflects behavioral intensity, and
- Total sales contribution, which reflects overall commercial importance
By plotting these together, the analysis avoids misleading conclusions that can arise when looking at revenue alone.
Insight:
High-revenue segments cluster around higher transaction frequency rather than unusually high spend per purchase. This indicates that category performance is driven primarily by repeat purchasing behavior, not one-time high-value transactions.
This view isolates purchase frequency to identify which customer groups exhibit habitual buying patterns.
Insight:
Older Families and Young Families demonstrate the highest purchase frequency, followed closely by Older Singles/Couples. These segments represent consistent, repeat-driven demand rather than volatile or episodic purchasing.
- Chip category performance is behavior-driven, not price-driven
- High-performing segments purchase frequently and predictably
- Low-frequency segments contribute less reliably to total sales, even when individual transactions are moderately priced
From a strategic perspective, this implies that:
- Increasing visibility and availability for high-frequency segments is likely to generate stronger returns than broad-based discounting
- Store layout and assortment decisions should prioritize segments with stable purchasing patterns rather than attempting to stimulate infrequent buyers
📓 store_trial_evaluation.ipynb
The second phase evaluates whether a store layout intervention led to measurable, incremental improvements in store performance, after accounting for natural variation and external trends.
The evaluation follows a structured causal framework:
-
Trial and Control Store Selection
Control stores are selected based on similarity in historical performance to ensure meaningful comparison. -
Pre-Trial Trend Validation
Performance metrics are compared before the intervention to confirm parallel trends across:- Total sales
- Number of customers
- Transactions per customer
-
Post-Trial Impact Assessment
Expected (no-trial) performance is estimated and compared to observed outcomes. -
Uncertainty Quantification
Confidence intervals are used to distinguish real uplift from random fluctuation.
While multiple trial stores were evaluated in the analysis, the README presents Store 77 as a representative example. This approach improves clarity while preserving analytical integrity.
All trial stores were assessed using the same framework, and conclusions are based on consistent patterns observed across stores, not a single data point.
This visualization demonstrates that the trial and control stores followed closely aligned trajectories prior to the intervention.
Interpretation:
The similarity in trends supports the assumption that post-trial divergence can reasonably be attributed to the layout change rather than pre-existing differences.
This chart quantifies incremental sales relative to the expected no-trial baseline.
Interpretation:
Post-intervention sales exceed expected values, indicating positive incremental impact associated with the layout change.
This visualization places observed sales within an uncertainty band to assess statistical significance.
Interpretation:
Actual performance exceeding the confidence interval strengthens confidence that the observed uplift reflects a genuine effect rather than random noise.
- Pre-trial alignment validates the control store selection
- Post-trial uplift is directionally consistent and statistically defensible
- Performance improvement is driven by behavioral changes rather than temporary customer spikes
Findings from both analytical phases were consolidated into an executive presentation (Quantium_analysis) structured using the Pyramid Principle.
Structure:
- Clear top-line recommendation
- Supporting insights
- Analytical evidence
This approach ensures stakeholders can quickly grasp conclusions while retaining access to analytical justification when required.
- Proceed with broader rollout of layout changes in stores with similar profiles
- Focus future optimization efforts on high-frequency customer segments
- Institutionalize trial–control evaluation for future interventions
- Avoid relying solely on aggregate sales metrics without behavioral decomposition
This project demonstrates analytical rigor suitable for retail strategy, commercial analytics, and data science roles. For questions or collaboration inquiries, please reach out via the repository contact information.