What data we actually have

DSA has sent us ten of the twenty-four datasets we asked for. The ten that arrived are rich enough to build on. The fourteen that didn't include two that quietly determine whether we can measure anything at all.

The estate splits cleanly in two. My Mai's extracts are eleven system pulls from JDE — shipments, item master, purchase orders, inventory, market data — delivered as flat files against a formal request specification. Brian's files are five Excel workbooks, and they are not extracts at all: they are the planning process. The SRX workbooks carry 79 sheets each, and the plan itself exists nowhere but inside them.

That asymmetry is the single most important structural fact in this document. The forecast we are being asked to improve is not a system output that can be queried. It is a spreadsheet built by hand, one channel at a time, from six manually entered drivers. Everything downstream of that — accuracy benchmarking, value estimation, automation — depends on reading it correctly first.

My's own request register — DSA_Data_Request_vShare.xlsx — lists 24 datasets with priorities, grains, join keys and DSA's own availability assessment. It is the most useful file in the estate, because it lets us measure delivery against intent rather than guess.

Five findings change what we do next. They are ordered by how much they affect the plan, not by how interesting they are.

This was the founding question: F01HistoricalForecast.xlsx does not agree with the B1 forecast in the SRX submission. Because if two sources disagree about what was forecast, "forecast accuracy" has no single definition and Step 2 cannot start.

It resolves cleanly. The two files disagree because they cover different scopes, not because either is wrong.

Seven business concepts appear in more than one place. Where two sources describe the same thing, we can measure whether they agree — and the answers range from a 0.1% tie to a sheet that disagrees with its own predecessor on two-thirds of its rows.

Duplication here is not waste. It is the only cross-check available, and in three cases it is the only route to data DSA told us did not exist.

Seven things exist in exactly one place. For these, "agreement" is not a question we can ask — which makes their quality assumptions load-bearing.

• F13 Golf Datatech and F14 NGF — external market data, unverifiable internally by design. Note the workbook's MARKET SIZE column is hand-keyed rather than drawn from F13, so those two could be cross-checked and currently aren't.
• T05 item master — the join spine. Nothing validates it, and seven of its attributes are empty.
• T06 channel dimension — delivered at customer grain with 40% of region labels missing.
• T07 pricing — 17% of items, no clubs. The workbook's cost tabs cover different ground and use a different structure, so they are a supplement, not a check.
• F07 BOM — single source and structurally void.
• The four non-Srixon brands — 24% of plan volume, visible only in F01, with no build-up behind them. This is the largest unverifiable block in the estate.

Sixteen data files under 01_Data/Client_data/, split by provenance. Sizes and row counts are measured, not reported.

Where each business concept authoritatively lives, and how a number travels from source system to submitted plan.

Item is the backbone, as expected. What matters is where the key formats diverge and where the grain shifts underneath a join.

79 sheets in the B1 Submitted file, 73 in the 2024 edition, 79 in the 2026 live file. They group into seven functional families.

Traced from the 2026 - SALES sheet by separating formula cells from literals. This is where human judgment enters the forecast — and therefore where both the automation opportunity and the accuracy opportunity live.

Every column in every delivered extract, with measured fill rates and cardinality. Fill rate is share of non-null rows.

So the numbers in Figure 6 can be reproduced or challenged:

• Plan comparisons match on the five-part key product code · product line · channel cat · sales type · channel, summing TOTAL. "Unchanged" means equal within a relative tolerance of 1e-6 on keys present in both files. Keys present in only one side are reported separately as dropped/added rather than counted as disagreement. Note the 2024-and-earlier sheets use SRI CAT/MODEL where later ones use DSP Product Code/Product Line; the comparison uses whichever pair the sheet carries.
• Shipment comparison aggregates T02 to product line × year × period after filtering to Srixon + Domestic Sales + Closed and converting quantity to pieces via PIECES_PER_UNIT from T05. Pieces-per-unit is available for 77.2% of T02 rows; unmatched rows are treated as 1:1, which understates the true total slightly.
• Grand Total rows in the workbook's pivot exports are excluded. Failing to do so double-counts every figure.
• Caveat on inventory. The four observations differ in filter scope — F04 covers two locations with no brand filter, the workbook tabs are pivot exports with their own filters. Treat the trend as indicative until scope is normalised.
• Not yet tested: the invoice-date sheet in 2025 forecast accuracy.xlsx (663,610 rows) against T02; the workbook's hand-keyed MARKET SIZE against F13; and DSA's own accuracy figure of 43.3% found in that workbook, which we should reproduce before proposing any alternative.

Answers we need before Step 2 can produce a defensible number.

The forecast-accuracy work that rests on this map lives in a separate document, B1 Forecast Baseline — the options catalogue for pairing a plan with actuals, the decisions taken, and the measured result for plan year 2025.

This atlas stays what it says on the front: a map of what data exists, how the pieces connect, where they agree, and what is missing.