# F1-dev-01 — Analytic firm dataset and revenue aggregates

## Question
Build a documented, analysis-ready version of the synthetic firm cross-section `data/firms.csv` and report: the number of distinct firms, mean and median revenue in millions of USD over firms with known revenue, and mean head count over all distinct firms.

## Data and exclusions
- Input: `data/firms.csv` (SHA-256 `78f119a9…625a`), 445 data rows (`n_input_rows`). Full audit in `out/state/data_audit.md`.
- Issues detected (`data_issues`):
  1. 45 rows repeat an earlier `firm_id`; every one is an exact full-row duplicate (no conflicting values). Handling: keep one row per firm. Result: 400 distinct firms (`n_unique_firms`).
  2. Mixed revenue units: some rows are in `USD_thousands`, the rest in `USD_millions`. Handling: divide `USD_thousands` values by 1000 (60 distinct firms). After conversion, revenue per employee for the two groups falls in the same range, so the `unit` labels look correct.
  3. 6 distinct firms have an empty revenue cell (`n_missing_revenue`): F0048, F0121, F0129, F0269, F0348, F0399. Handling: they are left out of the revenue aggregates but kept in `n_unique_firms` and `mean_employees`.
- Exclusion ledger (`out/exclusions.csv`): 445 → (drop exact duplicates, −45) 400 → (unit recoding, −0) 400 → (drop missing revenue, −6) 394 = `n_analysed`. The analytic dataset `out/analytic.csv` has 394 rows, one per firm.

## Method
This is deterministic data cleaning plus descriptive statistics (pandas). It uses no randomness (`random_seed` = null). Revenue aggregates are the arithmetic mean and median of `revenue_musd` over the 394 analysed firms. `mean_employees` is the mean head count over all 400 deduplicated firms. The baseline `mean_revenue_raw` is the mean of the raw `revenue` column over all 438 non-empty rows, taken as read, with no deduplication or unit conversion. `out/analysis.py` regenerates every output.

## Results
- Distinct firms: 400 (`n_unique_firms`). Firms with unknown revenue: 6 (`n_missing_revenue`). Firms analysed: 394 (`n_analysed`).
- Mean revenue: 249.96 million USD (`mean_revenue_musd`). Median revenue: 162.31 million USD (`median_revenue_musd`). The mean is above the median, so the distribution is right-skewed.
- Mean employees: 1101.4 (`mean_employees`).
- Raw baseline: 30,695.9 in the file's mixed units (`mean_revenue_raw`). This number has no meaningful unit. It is about 123 times the cleaned mean because the `USD_thousands` rows are mixed in without conversion, which shows why the unit recoding is necessary.
- Feasibility: complete (`feasibility`).

## Uncertainty
The estimands describe this file exactly; there is no sampling uncertainty about them. For reference only, if the 394 firms were treated as a random sample, the standard deviation of revenue would be 274.3 million USD (`sd_revenue_musd`) and the standard error of the mean 13.82 million USD (`se_mean_revenue_musd`). The results depend on the cleaning choices: whether to deduplicate, how to convert units, and that missing revenue is treated as missing completely at random (unknown values are not imputed).

## Limitations and open issues
- The data are simulated. Duplicates were identified by exact match on all columns. Near-duplicates with different `firm_id` values cannot be detected from this file.
- I assumed the `unit` column is correct on every row. A plausibility check of revenue per employee (85.9k–826.9k USD) found no rows that look mislabelled, but this cannot be fully verified.
- Excluding the 6 firms with missing revenue may bias the revenue aggregates if the missingness is related to firm size. Their head counts (326 to 3547) span the observed range.