top of page

Select a project to explore the interactive dashboard.

Loading insights may take a few seconds.

If a server error pops up, stay calm and refresh the page.

pCLV
Project Overview

Anticipating customer churn and forecasting future financial value is a cornerstone challenge in modern e-commerce. This project focuses on Predictive Customer Lifetime Value (pCLV) Modeling and Churn Simulation to pivot corporate strategy from reactive discounting to proactive revenue optimization. By implementing probabilistic machine learning models on transactional data—specifically capturing a cohort of 2.2K active retail customers representing $740K in expected near-term revenue—the project shifts the paradigm from reporting historical metrics to forecasting behavior over a 90-day horizon. The system provides an end-to-end analytical pipeline culminating in an executive-ready interactive simulator dashboard. This enables management to not only visualize current customer health and Pareto distributions but actively execute "What-If" operational scenarios to see how marginal improvements in retention or cross-selling impact the bottom line.

​

The Technical Pipeline

  • Python (Probabilistic Machine Learning & Financial DCF): Utilized the lifetimes ecosystem to deploy BG/NBD (Beta-Geometric/Negative Binomial Distribution) models to capture dynamic customer survival states and predict future purchase frequencies, alongside Gamma-Gamma models to estimate conditional expected Average Order Value (AOV). Incorporated financial continuous-time discounting (1% monthly rate) to calculate true Net Present Value (NPV) for a mathematically rigorous 90-day pCLV projection.

  • Tableau (Strategic Interactive Simulation & Executive UI): Engineered a web-app style What-If simulation engine leveraging dual-parameters (Retention Lift 0-5% and Upsell Lift 0-20%) to dynamically recalculate projected revenues. Overcame native Tableau layout restrictions by building a custom unified simulation matrix using synchronized multi-sheet layering, utilizing automated dynamic headers, conditional risk-based font coloring, and custom-bounded smooth red-green diverging heatmaps.

  • ​View python (pCLV Predictive Modeling) Code on Kaggle

​

Key Business Insight

  • High-Risk Revenue Salvage (Proactive Churn Mitigation): Identified a critical cohort of 88 "Fading Out" customers whose Probability Alive (P(Alive)) has plummeted below the 0.60 threshold, tying up $9K in high-risk baseline revenue. The simulator demonstrates strict logical integrity: by deploying a targeted 2% Retention Lift campaign without wasting margin on upsell incentives for at-risk users, the system isolates a clean, risk-adjusted salvage of $970, successfully pulling vulnerable accounts back into safety.

  • The Upsell Multiplier (The High-Velocity Growth Engine): The simulation matrix reveals that the company's highest financial leverage resides in driving order density among healthy clients. For the active Retail segment, achieving a 20% Upsell Lift expands the Expected AOV from $344 to over $412, unlocking an incremental $181K in projected future revenue (+20% top-line growth). This proves that driving cross-sell tracks for high-probability-alive clients yields immediate, massive financial returns.

  • Pareto-Driven Nurturing (Locking Down Core Value): The visual intelligence layer establishes that the top 20% of retail customers generate approximately 40% of total future revenue. By combining this distribution with the "Top Customer Explorer," marketing teams can seamlessly extract real-time lists of high-value VVIPs (such as Customer ID 17675 with a pCLV of $3,276 and 5 expected near-term purchases) to enroll them into premium, high-ROI nurturing tracks before competitor poaching occurs.

A/B Test
Project Overview

Measuring the true causal impact of advertising is one of the most complex challenges in digital marketing. This project focuses on Incrementality Testing for a "Summer Promotion" campaign to determine whether ad exposures genuinely drove new conversions, or if users would have converted anyway.

By analyzing a dataset of 588K total users, this project successfully quantified the campaign's exact incremental value, proving a +43% Relative Lift with a >99.9% statistical confidence level. The test confirmed that the campaign generated 4.34K incremental conversions and an estimated $434.30K in additional revenue. Beyond simply reporting the baseline metrics, this project dives deep into ad frequency distribution and time-series behavior to uncover critical budget inefficiencies and actionable optimization strategies.

​

The Technical Pipeline

  • Python (EDA & Feature Eng): Addressed extreme right-skewed ad frequency data using Logarithmic Binning to ensure statistical power across cohorts.

  • Statistical Testing (Z-Test): Applied a Two-Proportion Z-Test to validate conversion rate differences, calculating Absolute/Relative Lift and confirming a p-value < 0.001 (>99.9% confidence).

  • Power BI (Advanced BI & UI): Built an executive-ready dashboard utilizing advanced DAX (CALCULATE, ALL), conditional data bars, and dynamic layering.

  • ​View Python (Z-test) Code on Kaggle

​

Key Business Insight

  • Cap Frequency at 100: Stop wasting budget on outliers (>100 ads). Reallocate spend to push the under-exposed majority into the 61-100 sweet spot (+7.15% peak lift).

  • Early-Week Focus: Shift inefficient Thursday and weekend budgets to aggressively target Mon-Wed (Tuesday peaks at +1.60%).

  • Prime Time Bidding: Increase bid multipliers during daily conversion spikes at 05:00-07:00 and 20:00-22:00.

Sentiment
Project Overview

Unstructured customer feedback contains the most critical signals for product optimization and brand survival, yet parsing thousands of app store reviews remains a notorious bottleneck. This project builds a production-grade Automated Text-Mining & UX Diagnostics Pipeline designed to ingest, classify, and visualize user feedback for Hong Kong’s leading e-commerce platform, HKTVmall.

​

By scraping App Store reviews via Python API, deploying the Gemini Large Language Model (LLM) for multi-label semantic classification, and engineering an executive-ready Tableau dashboard, this project successfully transformed raw, chaotic user complaints into actionable product engineering inputs. The analysis diagnosed a severe drop in user sentiment (Avg Rating: 1.68) across recent app deployments, successfully isolating hidden UI/UX friction points, tracking version-specific performance degradation, and uncovering non-systemic brand equity risks to empower cross-departmental PR and product alignment.

​

The Technical Pipeline

  • Python API (Automated Data Ingestion): Developed an automated script leveraging the Apple App Store scraper API to fetch app reviews, consolidating metadata including review content, historical timestamps, ratings, and application versions into a clean, unified data stream.

  • Gemini LLM & Google Sheets (Intelligent Semantic Categorization): Built a serverless text-classification workflow utilizing Gemini via Google Sheets LLM functions. Implemented zero-shot prompt engineering to clean noisy text and precisely map unstructured Cantonese/English reviews into 6 functional problem domains (e.g., Performance, Ad Interference, Logistics) and across 4 distinct user journey stages. Resolved complex many-to-many data dependencies where a single system crash dynamically generated friction across multiple journey touchpoints.

  • Tableau (Executive Visual Intelligence & UX Diagnostics): Engineered a high-density visual utilizing custom color semantics (separating pure crisis metrics from actionable salvageable pools). The dashboard integrates a Friction Distribution Across Funnel horizontal bar chart to immediately spotlight process choke points, a Version Impact Heatmap to audit technical stability from v5.4.0 to v5.10.0, and a dedicated Brand Equity Risk Module to serve as an early-warning system for executive management.

​

Key Business Insight

  • The Product View Choke Point (Critical Funnel Friction): The analysis revealed that the Product View stage is our single largest bottleneck, harboring a staggering 96 friction incidents—nearly double any other funnel stage. Cross-referencing this with the Version Impact Heatmap proves that this isn't a generic issue, but a severe performance regression (system lag, abrupt app crashes) introduced heavily in version 5.5.0 and returning in v5.10.0. Engineering teams must immediately prioritize stability patches for these specific build versions to unlock the conversion funnel.

  • Monetization vs. Retention Tension (Ad Interference): Ad Interference emerged as a primary secondary driver of negative sentiment, accounting for 25 high-severity reviews. Users are actively penalizing the core app experience due to aggressive pop-ups and ad placement. Product Management needs to recalibrate the ad-display frequency immediately to balance monetization goals against user retention, as this friction is heavily concentrated during the crucial Payment phase (34 friction points).

  • Brand Equity & PR Vulnerabilities (Non-Systemic Risks): Beyond technical glitches, the dashboard exposed 3 distinct non-systemic reputation threats silently eroding long-term brand equity: 12 reviews weaponized ethical concerns regarding Animal Testing, 10 reviews cited aggressive competitor platform advantages, and 4 reviews explicitly attacked product quality via "Taobao" sourcing perceptions. A proactive corporate defense strategy requires immediate alignment with PR for public messaging and the Merchandising team to stringently vet supplier origins.

RFM
Project Overview

Understanding customer behavior and maximizing lifetime value is one of the most critical challenges in e-commerce. This project focuses on Behavioral Customer Segmentation using an RFM (Recency, Frequency, Monetary) framework to eliminate inefficient, "one-size-fits-all" marketing. By analyzing a dataset of 93K total users representing $15.4M in revenue, this project successfully classified customers into 4 distinct behavioral cohorts using K-Means machine learning. The system provides an automated pipeline from data warehouse to an executive-ready dashboard, giving management immediate visibility into core segments. Beyond simply reporting baseline metrics, this project dives deep into segment-specific purchasing habits and revenue contribution to uncover targeted win-back strategies and high-ROI nurturing tracks.

​

The Technical Pipeline

  • BigQuery SQL (Data Consolidation): Consolidated large-scale relational transactional databases by query-optimizing multi-table joins (orders, customers, payments, and reviews) into a unified, high-integrity core dataset.

  • Python (Feature Engineering & ML Clustering): Engineered customer-level RFM metrics, resolved high right-skewness via log transformation, and scaled features using StandardScaler; utilized the Elbow Method to determine an optimal K=4 and executed K-Means to precisely segment the 92K customer base.

  • Tableau (Executive Visual Intelligence & UI): Built a web-app style dashboard utilizing custom "Master UI Buttons" for foolproof filtering, a noise-free Top 10 category tree-map with blue-green diverging heatmaps, and strict action controls.

  • ​View Python (K-Means Clustering) Code on Kaggle

​

Key Business Insight

  • VVIP Retention (Preventing Royalty Fade): Urgent intervention is required for our 2.8K VVIPs. While they are our most loyal segment with the highest purchase frequency (2x) and a premium average spend of $316, their average recency has dilated to a dangerous 220 days. This elite group is actively slipping into dormancy. We must immediately launch exclusive VIP loyalty perks, experiential rewards, or concierge outreach to rebuild their sense of belonging and trigger their next purchase before they churn.

  • Potential VIP (The Growth Engine): Prioritize the 27.6K high-value customers who serve as our primary revenue driver, contributing $8.8M (57% of total revenue) with a remarkable average order value of $320. Implementing targeted cross-selling or milestone incentives here offers the highest financial leverage for the company.

  • Targeted Home Win-Back (Churned): Re-engage the 27.2K severely lapsed customers who have been inactive for an average of 424 days, yet still represent $3.2M in historical value. To efficiently capture this low-hanging fruit, deploy hyper-focused reactivation campaigns tailored specifically around their preferred "bed bath table" and "health beauty" categories.

Churn Probability
Project Overview

This end-to-end predictive analytics project transitions complex subscriber data into robust, statistically validated commercial strategies. By deploying a rigorous machine learning pipeline integrated with an interactive executive dashboard, this solution empowers telecom leadership to proactively mitigate churn, optimize marketing spend, and quantify financial risk in real time.

​

The Technical Pipeline

  • Robust Data Engineering & Modeling: Extracted and preprocessed subscriber features using SQL. Implemented a 5-Fold Cross-Validation and Out-of-Fold (OOF) simulation framework in Python (XGBoost) to generate 100% leak-free, mathematically honest customer churn probabilities across the entire population.

  • Asymmetric Cost & Decision Threshold Optimization: Rejected the naive 0.5 default classification baseline. Engineered a Threshold Optimization loop to maximize the F1-Score, aligning the model's decision boundary with commercial reality—effectively balancing the high cost of customer defection (False Negatives) against the low cost of proactive retention incentives (False Positives).

  • Advanced DAX Architecture: Translated the optimized model outputs into dynamic business metrics within Power BI, notably Revenue at Risk, which instantly captures the monetary impact of potential churn based on real-time risk scores.

  • Strategic UI/UX Design: Formulated a premium, executive-level dark mode interface. Implemented advanced conditional formatting rules (dynamic color-coded risk alerts synced with the optimized threshold) and customized single-direction visual filtering to ensure a distraction-free, high-performance data drilling experience.

  • ​View Python (XGBoost + 5-Fold OOF) ML Code on Kaggle

​

Key Business Insight

  • The Macro Drivers (Feature Importance): Feature importance analysis revealed that Month-to-Month Contracts and Fiber Optic subscription anomalies are the primary macro-drivers of customer defection, outranking pure pricing factors.

  • The "Tech Support" Lever: At the operational level, the dashboard uncovers that customers without technical support represent the highest immediate concentration of churn risk, accounting for $39K in Revenue at Risk.

  • Actionable Business Impact: Executives can utilize the top-level parameters to simulate tactical interventions in real time. For example, converting Month-to-Month users to long-term contracts or deploying targeted tech support instantly visualizes the simulated mitigation of risk, proving how data-driven operational decisions directly protect the company's bottom line.

bottom of page