A boutique hotel group in Thailand has three data systems that don't talk to each other. You combine booking data, guest reviews, and revenue reports into one analysis that shows which property differences are real and what drives guest satisfaction.
The discipline skills: joining datasets with DuckDB and pandas, running ANOVA and pairwise tests with assumption checking, computing effect sizes alongside p-values, building a prediction model with regularization and cross-validation, and communicating findings that separate statistical significance from practical importance.
The AI-direction lesson: AI picks statistical tests without checking whether the data meets the test's assumptions. It runs parametric tests on non-normal data, skips multiple comparison corrections, and reports p-values without effect sizes. The skill this time is catching AI's default test selection — recognizing when a t-test should be a Mann-Whitney, when a battery of pairwise tests needs a Bonferroni correction, and when "p < 0.05" is not the whole answer.
Your Role
You combine three sources into something that tells a fair story. There is no provided answer key this time. When you join datasets and run a statistical test, you verify the work yourself. A second AI reviews your methodology from a fresh perspective.
What's New
Last time, you built a prediction model on a single, familiar dataset. Verification targets told you what honest results looked like. You caught data leakage by checking against those targets.
This time, the data starts messy. Three sources that were never designed to work together. Different date formats, different naming conventions, missing values. Before any analysis can run, you need to decide how to combine them — and those decisions affect every number downstream. This transition from clean datasets to messy multi-source reality reflects what data science actually looks like in practice — working data scientists spend roughly 30-50% of their time on data work (acquisition, cleaning, joining, profiling), and only 10-15% on actual modeling.
The analytical work is also new. You will test whether differences across properties are statistically significant, measure how large those differences actually are, check whether the statistical tests are even valid for this data, and build a model to identify satisfaction drivers. A p-value without an effect size is incomplete. A test run on data that violates its assumptions produces a number that looks real and is not.
Tools
- Python 3.11+ via your conda "ds" environment
- Jupyter Notebook for the analysis
- pandas for data handling
- DuckDB for SQL-based multi-source analysis (new this project)
- scikit-learn for prediction modeling (Ridge, Lasso, cross-validation)
- scipy for statistical tests and assumption checking
- scikit-posthocs for post-hoc pairwise tests (Dunn's test after Kruskal-Wallis)
- matplotlib / seaborn for visualization
- Claude Code as the AI you direct
- Git / GitHub for version control
Materials
You receive:
- Three datasets: bookings, guest reviews, and monthly revenue reports
- A data dictionary for each source
- A methodology memo template that you fill in as you work
- A project governance file (CLAUDE.md) for Claude Code
- Somchai's email explaining what he needs
No verification targets. No step-by-step pipeline. The methodology memo gives you a structure to record your decisions. The data dictionaries tell you what each column means. You direct the work and verify it yourself.