I turn messy data
into decisions people
can act on.

Business problem

A single national pricing strategy was a guess, not a decision. The brand had no chain-specific read on how price actually moved demand.

Approach

Worked with 4 years of Nielsen scanner data (6 brands, 7 supermarket chains) to specify, estimate and validate a market-response model. Cleaned mislabeled promotion prices, used multiple imputation (MICE) for missing values, and specified a multiplicative log-log demand model with competitor price tiers, chain fixed effects, seasonality and an AR(1) term.

A Chow test (F = 5.43, p < 0.001) showed pooling all chains into one model was statistically wrong. So I re-estimated chain-by-chain instead. Ran full diagnostics (VIF/GVIF, Breusch-Pagan, Goldfeld-Quandt, Breusch-Godfrey) and corrected inference with Newey-West HAC standard errors.

Outcome

Long-run price elasticity ranged from -0.45 (loyalty-driven channel) to -3.24 (highly price-sensitive channel) A model directly usable for chain-by-chain pricing and promotion planning. Out-of-sample MAPE was 17.7% overall (7.8% for the best chain).

Business problem

Operating in a liberalized energy market with near-50% annual churn, the client had no systematic way to flag at-risk customers.

Approach

Formulated and statistically tested five hypotheses about churn drivers using t-tests and chi-square tests. Engineered features including a flexible-contract flag, winsorized outliers, and encoded categorical predictors. Built and benchmarked six models. Logistic regression, CART, Random Forest (800 trees), SVM, and a neural network. Evaluated on hit-rate, top-decile lift, and Gini.

Outcome

Customers on flexible, no-penalty contracts were nearly 7x more likely to churn The strongest single driver. Counter to the initial hypothesis, high energy usage predicted churn, not loyalty. Delivered a concrete retention playbook: contract-lock incentives, bill-smoothing, human onboarding calls, and insulation subsidies.

Business problem

NYC taxi drivers have no data-driven way to decide where to start a shift or which trips to prioritize.

Approach

Working from 100,000 trips enriched with demographic, economic and weather data, ran a two-part tipping model (logistic + OLS on log-tip amount) and a shift-optimization model (Poisson, extended to Negative Binomial after diagnosing overdispersion). Rather than ranking zones by average fare. Which rewards drivers stuck in profitable-looking gridlock. Engineered a "net revenue per minute" metric and validated it against mechanical bias.

Outcome

Doubling the fare raises tipping odds by 28.6% and tip size by 94%. Manhattan trips show 2 to 4x higher tipping odds than Bronx/Queens. Delivered borough- and hour-level recommendations drivers can act on directly.

Business problem

Sol de Janeiro needed evidence-based guidance on which influencers and content styles actually drive engagement, rather than relying on follower-count intuition.

Approach

Built a 250-post dataset across 5 influencers, extracting sentiment/subjectivity via VADER and TextBlob, structural caption features, and CLIP-based image classification (casual vs. luxury style, brightness/contrast/saturation). Built OLS regressions on log(likes) and log(comments), then tested a casual-style × emoji-count interaction.

Outcome

CLIP "Casual" was the only image feature significant for likes and the strongest for comments. Emojis boosted comments significantly. But only on posts already casual. Recommended shifting budget toward authenticity-over-production creators.

Business problem

Wehkamp ran a 22-day advent-calendar loyalty campaign across 16,035 participants and needed a rigorous read on whether it worked. From messy, multi-source logging data.

Approach

Diagnosed and fixed four distinct data-quality issues: mismatched ID headers causing ~15% join failures, UTC vs. Amsterdam timezone misalignment, inconsistent customer-ID formats, and two extreme outlier customers skewing demand figures. Ran McNemar's test to confirm the order-rate lift against a prior-year baseline.

Outcome

Confirmed +7.8pp order-rate uplift (38.8% → 46.6%), 84.75% overall open rate. Identified that "Tip" reward cards. Which carried no tangible reward. Caused a reproducible ~36% single-day engagement drop. Flagged the missing control group and margin data rather than overstating the ROI case.

Business problem

Kruidvat was losing market share specifically in Groningen despite running promotions. Intratuin wanted to expand into Germany without diluting its Dutch brand identity.

Approach

For Kruidvat, decomposed sales using the penetration × frequency × spend framework and isolated declining household penetration as the actual driver. Not weak promotions. For Intratuin, applied Hofstede's cultural dimensions to redesign the customer journey for German expectations, and compared two candidate states using population density, income, and competitive saturation.

Outcome

Household penetration had declined from 85.8% to 84.1% The real root cause. Designed a targeted 0.5% automatic cashback mechanism projected to lift purchase frequency by +20%. Delivered a complete go-to-market package for Intratuin's German entry.

Business problem

Linckr wanted to launch a digital e-wallet but had no clear view on which sector to target first, risking spreading a limited launch budget too thin.

Approach

Interviewed stakeholders to understand internal priorities and constraints, then ran market research on candidate sectors and cross-checked findings against academic literature on payment-adoption and market-entry sequencing. Rather than relying on stakeholder opinion alone. Built the sector-prioritization strategy: assessing each sector's readiness against academic frameworks, then translating that into a go-to-market sequence.

Outcome

Recommended education and real estate as priority launch sectors, backed by a research-grounded rationale, plus a sector-specific positioning approach for each.

Business problem

The brief: identify and pitch a viable brand extension for an existing consumer brand.

Approach

Built the case around a specific consumer pain point. Female gym-goers' discomfort in male-dominated fitness spaces. And proposed extending Victoria's Secret into a women-only athleisure studio, leveraging the brand's existing equity in body confidence. My role covered three components: I designed the membership pricing structure, wrote the pitch video script, and conducted the market research underpinning the concept's demand rationale.

Outcome

Selected as a top project in the course.