Analytics · Analytics Showdown 2025 — Santa Clara University
Nazava × Shopee Sales Optimization
Forecasting models that won 1st Prize by translating analytics into promotion strategy
The Challenge
Nazava Water Filters lacked data-driven promotion strategy on Shopee marketplace. Unclear which timing, discounts, or campaigns actually drove sales. Competing with 1,000+ sellers in the same category, with no forecasting tools and decisions made on gut feel. Led to stockouts, excess inventory, and missed revenue opportunities.
My Role & Approach
Owned the predictive modeling pipeline, Tableau dashboard design, and final presentation to a panel of 6 judges and 100+ audience. Made the strategic decision to lead with business impact rather than model accuracy in our presentation — framing every metric in terms of revenue gained or cost saved rather than statistical measures.
Product Thinking
Problem Framing
Most teams at the competition led with model accuracy. We reframed: Nazava doesn't care about R² — they care about 'when should I run my next Shopee promotion and at what discount?' This made us build a decision tool, not a forecasting model.
Key Tradeoffs
XGBoost vs. ARIMA: ARIMA was simpler to explain but couldn't capture flash sale non-linearity (62% accuracy). We chose XGBoost knowing it's harder to interpret, then invested the saved accuracy into better business storytelling. Also chose 3 high-impact insights over 10 mediocre ones for the final presentation.
What We Didn't Build
Scoped out real-time pricing optimization, competitor price scraping, and automated Shopee ad bidding. Each would have diluted our core message. The competition was 24 hours — ruthless prioritization was the differentiator.
Cross-Functional Leadership
Led a 3-person team. Owned the modeling pipeline and presentation strategy. Coordinated with teammates on data cleaning and visualization. Presented findings to 6 industry judges and 100+ audience members.
Solution & Execution
Predictive demand model (XGBoost) that captured non-linear patterns from flash sales better than traditional ARIMA. Customer segmentation (K-means clustering) for targeted marketing. Pricing elasticity analysis with actionable discount recommendations. Real-time Tableau dashboard for KPI tracking. Delivered a complete playbook for Shopee campaign optimization.
Tech: Python, Tableau, Scikit-learn, Pandas, XGBoost, K-means
Impact
- 87% improvement in promotion decision speed — from weeks of analysis to hours
- 1st Prize — Analytics Showdown 2025 at Santa Clara University
- 84% demand forecast accuracy using XGBoost (vs 62% with ARIMA baseline)
- Presented insights to 100+ attendees and 6 industry judges
Challenges & Pivots
Challenge: First draft of our presentation was too technical — model metrics without business context
Resolution: Completely reframed around business storytelling: every metric translated to revenue impact. Instead of 'R² = 0.84', we said '84% accuracy means Nazava can reduce stockouts by X%. This approach won the competition.
Challenge: Shopee sales data had extreme non-linearity from flash sales events that broke ARIMA models
Resolution: Switched to XGBoost which captured interaction effects between promotions, seasonality, and competitor pricing. Accuracy jumped from 62% to 84%.
“Your team's ability to translate complex forecasting models into actionable business recommendations stood out. The presentation demonstrated mature product thinking.”
— Judge Panel, Analytics Showdown 2025
What I Learned
- The best analytics bridges statistical rigor with business storytelling — judges cared about revenue impact, not R² scores.
- Competition environments force rapid prioritization — we focused on 3 high-impact insights rather than 10 mediocre ones.
- Forecasting models need business context: a technically perfect model that doesn't drive decisions is useless.
Next case study
Post-Release Optimizer