In most analytics work, EDA is a private warm-up. Here it is a contract with the client and the foundation of the model itself, for three reasons:

• The client signs off on the EDA's outputs. Data Validation Reports (DVRs) are presented to the client to lock alignment on the datasets used for modelling; only on client approval do the individual reports consolidate into the final Analytical Data Set (ADS). A dispute about data quality after modelling starts is a dispute you already lost.
• The bounds are set from the EDA. Modules 1.6 and 1.7 will show you that every coefficient bound starts from numbers computed at this stage - base-variable correlations and per-channel ASSR. A wrong EDA does not just embarrass you; it quietly poisons every bound downstream.
• The 8-week Build clock is running. An anomaly caught in EDA costs an email. The same anomaly caught in model iteration costs re-runs inside a committed window.

What the program's EDA covers, per the pipeline's first stage: process data to business specifications; clean and flag anomalies - missing values, duplicates, outliers; generate validation reports with visualisations and summary statistics; and run correlation analysis across channels to assess media impact against sales. In practice the checks from module 1.2 run here too: hierarchy sum-consistency, time-frame alignment across sources, the TDP decomposition identity, and a diff against the previous raw file when one exists.

The methodology workbook's Data EDA sheet demonstrates the working pattern on a ~36-month example ADS slice (revenue, price, distribution, spend, CPI, competitor TV spend). Two moves matter:

Move 1 - every variable gets typed, and each type gets its own validation metric. You do not stare at 25 columns with the same lens; you ask each one the question appropriate to its role:

Move 2 - compute per-channel ASSR before any model exists. ASSR (ad spend to sales ratio) is each channel's spend divided by total revenue. It needs no model, only the ADS - which is exactly why it is the program's earliest sanity read. It ranks channels by weight, exposes a channel whose spend share is wildly out of line with its claimed importance, and becomes the benchmark that media contributions are later set against (module 1.7). Alongside it, the base-variable correlations (distribution vs sales, price vs sales, CPI vs sales) anchor the base-side contribution expectations (module 1.6).

The sheet closes with three standard visual checks - Spend vs Revenue, Distribution vs Revenue, Price vs Revenue line charts - the plots you eyeball before believing any correlation number. And when reading a decline, follow the KT order: volume vs price first; if that does not explain it, distribution; then promo; then media last - media typically explains only ~12-20% of volume, so it is rarely the headline of a sales story.

Every cell's Data sheet carries a data_summary field - one narrative line that tells a reader (often a delivery lead scanning 80 cells) where this cell's data actually stands. The program style packs four things into one line: what is resolved, what is being chased, who owes it and by when, and any granularity caveat. Compare (all synthetic):