Define the extract and build a useful data profile

Understand the method

Exploratory data analysis begins with a question and a deliberate look at the data. For this module, Northstar Shop asks whether May’s larger completed-order sales justify expanding a sales initiative. Before comparing numbers, establish what each row means, which records belong in the comparison and what the extract cannot tell you. A polished chart cannot repair a population that changes between calculations.

A new period in the same fictional business

Modules 1–3 used February and March. Module 4 introduces a separate April–May 2026 teaching extract. It does not revise earlier results or describe real customers. The verified file contains 44 unique orders, 22 in each month. Each month has 20 completed, one returned and one cancelled order. The raw file contains 46 records because two identifiers occur twice. One repetition is identical; the other has conflicting channel values. A row count and an order count therefore answer different questions.

Write the analytical contract

The business measure is the recorded value of completed orders, in EUR. These are order values, not audited revenue, cash collected, profit or customer lifetime value. For the headline calculation, include completed orders dated from 1 April through 31 May. Use exclusive upper boundaries of 1 May and 1 June for monthly filters. Keep returned and cancelled rows available for a separate status analysis. The business-key grain is one verified order per order_id.

The supplied dictionary defines delivery_days as a recorded elapsed duration for a completed order. Blank means no measurement was supplied. It does not mean zero days. There is no customer identifier, advertising spend, product margin, transaction timestamp or experiment assignment. You cannot calculate customer conversion, profitability, daily operating speed or marketing causality from these fields. Listing these absences early prevents attractive but unsupported metrics later.

Build a compact profile

Start with total records, distinct identifiers, duplicate groups, date parsing, category frequencies, numeric ranges and blank counts. Profile the untouched raw file first. Count blank order values separately from negative values and count invalid delivery text separately from missing delivery measurements. A single “bad rows” total hides different actions and can double-count a record with several issues. Preserve record identifiers beside every exception.

After preparation, repeat the profile on the verified population and reconcile row counts. Record both observed date bounds and the intended reporting window. The last order is dated the 22nd in each supplied month; that does not prove the business had no orders afterwards. For the exercise we treat the fixture as the complete supplied extract, not independent proof of complete company activity. The stated comparison is descriptive of these records.

Make the profile useful

A useful profile connects a finding to a consequence. “A duplicate key exists” leads to checking whether it is an export repetition or a conflicting business record. “There is a blank duration” leads to reporting observed coverage before comparing averages. Keep the raw extract unchanged and maintain your decisions separately. This makes your exploration reproducible rather than a sequence of undocumented edits.