The Brief
Hassan El-Amin founded Nile Compass Tours in Cairo six years ago. Private cultural tours, multi-day Egypt itineraries, Nile cruise packages. He started with one van and a phone. Now he runs 20 staff, a network of freelance guides, and about 3,000 bookings a year.
The growth is the problem. Three years ago: 1,200 bookings. Two years ago: 2,100. Last year: 3,000. Hassan does not know why. Eighteen months ago he shifted his marketing budget from print ads in travel magazines to Instagram influencer campaigns and Google Ads. Around the same time, the Egyptian pound weakened, making Egypt cheaper for European visitors. A year ago he launched premium Luxor packages. Any of those could be driving growth. All of them could. None of them could.
He is making investment decisions -- hiring guides, expanding to Aswan, considering a Nile cruise partnership -- and he is doing it on instinct. He has the data. He wants evidence. His silent partner, an accountant, wants methodology.
Your Role
You deliver an inferential analysis of Hassan's booking data. But before you can analyze anything, you need to answer a question nobody is asking for you: what kind of question is Hassan actually posing? "I want to understand our booking patterns" is not a question type. It could mean description, inference, prediction, or causation. Each framing produces a completely different analysis. The methodology depends on the answer.
This is the first time nobody tells you the question type. The analysis specification, the validation approach, the methodology -- all of it follows from a decision you make and justify.
What's New
Last time, you built a demand forecasting model and discovered that preparation decisions -- temporal splitting, leakage prevention, feature timing -- determine whether a model produces honest predictions or impressive-looking numbers. You owned the preparation design.
This time, you own the analytical framing. When you ask AI to plan the analysis, it will propose something technically impressive. Whether that serves Hassan's actual decision is a judgment call that belongs to you. You will also design the validation strategy from scratch -- deciding what to check and why, not following a provided checklist. Meta-prompting enters: using AI to help you figure out what verification you need for territory you have not navigated before.
The hard part is recognizing that AI's confident methodology suggestion is a bias, not an answer.
Tools
- Python 3.11+ via your conda "ds" environment
- Jupyter Notebook for the analysis
- pandas for data handling
- statsmodels for OLS regression, hypothesis tests, and assumption checks
- scipy for statistical tests and effect size calculations
- scikit-learn for supplementary modeling if needed
- matplotlib / seaborn for visualization
- Claude Code as the AI you direct
- Git / GitHub for version control
Materials
You receive:
- Two datasets: three years of booking data (~6,300 rows including cancellations) and monthly marketing spend by channel (144 rows)
- A data dictionary explaining both datasets, including a note about self-reported marketing attribution
- A methodology memo template with new sections for question type determination and validation strategy documentation
- A project governance file (CLAUDE.md) for Claude Code
- Hassan's email explaining what he needs