You're evaluating whether a pricing change worked for a spice exporter in Galle, Sri Lanka — determining if three months of data can answer the question, whether the detected effect is large enough to matter, and whether keeping the new pricing makes financial sense.
The discipline skills: power analysis (whether the data can detect the effect you're looking for), distinguishing statistical significance from practical significance, computing ROI for an experimental result, and communicating an honest answer when the conclusion might be "we can't tell yet."
The AI-direction lesson: directing AI through layered judgment. AI computes a p-value without computing power — so "no significant difference" could mean no effect or insufficient data, and AI won't tell you which. AI reports "the test is significant" without evaluating whether the effect size justifies the cost of implementation. Each layer — power analysis, test selection, practical significance, ROI — requires a separate directing act with explicit constraints. AI doesn't stack these layers itself. You build the argument one directed step at a time.
Your Role
You're evaluating a business decision with data. The pricing change happened. The question is whether it worked -- and "worked" means different things depending on which metric you choose. Revenue can increase while profit stays flat if the discounts eat the margin. Your first job is to define what success means before you run any test.
Then: can the data actually answer the question? Three months of European orders may not be enough to detect a meaningful effect. You'll need to determine whether you have enough data before you interpret the result.
What's New
Last time, you investigated an ambiguous business complaint and framed your own analytical questions for the first time. You found four hidden problems in a subscription business and delivered audience-adapted findings.
This time, the analytical question is clearer -- did the pricing change work? -- but the answer is harder. You'll encounter power analysis for the first time: the forward-looking question of whether your data can detect the effect you're looking for. And you'll learn to distinguish statistical significance from practical significance -- a real effect that costs more to implement than it generates is a finding, not a success.
Tools
- Claude Code (with DuckDB MCP active)
- DuckDB (local database, MCP-connected)
- Python (pandas, scipy.stats, statsmodels, matplotlib/seaborn)
- Jupyter notebooks
- Git and GitHub
- CLAUDE.md (project memory -- established practice)
Materials
- Transaction data -- two years of spice export orders across three markets.
- Buyer data -- buyer-level records with market segment and relationship history.
- Cost data -- monthly cost breakdown by product, including seasonal variation.
- Pricing structure -- old per-kilo rates and new tiered volume discount structure.
- CLAUDE.md -- project governance file with empty sections for metric definitions, investigation framework, and analytical constraints. You populate these as the analysis unfolds.