DenyTranDFW's picture
Add Dockerfile, docker-compose, setup script, SQL stored procedures, schema docs, and indexes
af136d8 verified
# `sec_fsnds` — Schema Reference
A compact, normalized SQL Server database of the SEC Financial Statement and
Notes Data Sets (FSNDS). Every 10-K, 10-Q, and similar filing that appears in
the SEC's quarterly data set dumps is loaded here, with the raw text columns
normalized into lookup tables so the common query paths (filings → tags →
numeric facts) stay small and cache-friendly.
---
## Database properties
| Property | Value |
|-------------------|-----------------------------|
| Server version | SQL Server 2022 (Linux) |
| Compat level | 160 |
| Collation | `Latin1_General_CS_AS` |
| Recovery model | FULL |
| Data file size | ~300 GB (~251 GB used) |
| Log file size | ~1 GB (nearly empty) |
Collation is **case-sensitive**. `Assets` and `assets` are different tag names.
All SP parameters and identifiers in this package respect that.
---
## Tables at a glance
17 tables total: 9 core FSNDS tables, 7 lookup tables, and one provenance
table.
| Table | Rows | Data size | Notes |
|----------------|-----------|-----------|-------------------------------------------|
| `num` | 388 M | ~66 GB | Numeric XBRL facts |
| `pre` | 269 M | ~57 GB | Presentation order and labels |
| `txt` | 48 M | ~100 GB | Narrative (text) XBRL facts |
| `tag` | 43 M | ~8 GB | Distinct tag identities (tag + version) |
| `tag_text` | 43 M | ~10 GB | Long-form label and doc per tag |
| `cal` | 37 M | ~2 GB | Calculation linkbase (parent/child arcs) |
| `ren` | 24 M | ~6 GB | Rendering / report structure |
| `sub` | 820 K | ~320 MB | Submissions (one row per filing) |
| `dim` | 5.9 M | ~74 MB | Dimensional context IDs (stripped) |
| `lkp_version` | 495 K | ~34 MB | Namespace + version string → `versionID` |
| `lkp_coreg` | 54 K | ~3 MB | Co-registrant strings → `coregID` |
| `lkp_uom` | 16 K | ~1 MB | Unit of measure string → `uomID` |
| `lkp_datatype` | 124 | <1 MB | XBRL datatype string → `datatypeID` |
| `lkp_prole` | 34 | <1 MB | Preferred-role label → `proleID` |
| `lkp_stmt` | 8 | <1 MB | Statement code (`BS`, `IS`, `CF`, …) |
| `lkp_lang` | 1 | <1 MB | Language code (typically `en-US`) |
| `fsnds` | 75 | <1 MB | Source-file provenance (one per TSV file) |
---
## Core table shapes
### `sub` — Submissions (filings)
One row per filing. This is the hub of the database; almost everything joins
back to `subID`. The fields mirror the SEC's `sub.tsv` dump with one added
`subID` primary key.
Key columns:
- `subID` (PK, `int` identity)
- `adsh` (`char(20)`, unique) — the SEC accession number, e.g.
`0000320193-23-000106`
- `cik` (`int`) — filer CIK
- `name` (`nvarchar(250)`) — filer name at time of filing
- `sic` (`smallint`, nullable) — Standard Industrial Classification code
- `form` (`varchar(10)`) — `10-K`, `10-Q`, `8-K`, `20-F`, etc.
- `period` (`char(8)`, `yyyymmdd`) — period of report
- `fy` (`smallint`), `fp` (`char(2)`) — fiscal year / fiscal period
- `filed` (`date`), `accepted` (`datetime2`)
- plus ~30 address / filer-metadata columns
### `num` — Numeric facts
One row per reported numeric value. This is the biggest hot table by row
count.
- `numID` (PK, `bigint` identity)
- `subID`, `tagID`, `versionID`
- `ddate` (`date`) — reporting period end date
- `qtrs` (`tinyint`) — `0` = balance-sheet point-in-time; `1` = 3-month flow;
`4` = annual flow; other values appear for non-standard durations
- `uomID` — unit (USD, shares, pure, etc.)
- `value` (`decimal(28,4)`) — the numeric value
- `dimID`, `dimn``dimn = 0` means no dimensions applied; `dimn > 0` rows
are segmented breakdowns keyed by `dimID`
- `iprx` (`tinyint`) — "instance primary flag", SEC source attribute.
Canonical rows use `iprx = 0`
- `coregID` — non-null means the row belongs to a co-registrant and should
usually be excluded from consolidated analysis
### `pre` — Presentation
- `preID` (PK), `subID`, `stmtID` (`BS`/`IS`/`CF`/…), `report`, `line`, `tagID`
- `plabel` — the human-readable line label as rendered on the filing
- `negating` — flip the sign before displaying (e.g. "less: allowances")
- `inpth` — "in parenthesis" marker; typically skipped when building a
statement view
Join `pre` to `num` on `(subID, tagID)` to go from "line 12 of the income
statement" to the actual dollar amount, and order by `(report, line)` for
presentation order.
### `tag` / `tag_text`
- `tag` — one row per distinct `(tag, versionID)` pair. Includes custom
tags; `custom = 1` marks filer extension taxonomies.
- `tag_text` — `tlabel` (short label) and `doc` (full concept definition);
optional, split out to keep `tag` narrow.
### `txt` — Narrative facts
Similar shape to `num`, but holds the `nvarchar(max)` values that correspond
to textBlockItemType XBRL tags — accounting policies, risk factors,
commitments, and everything else non-numeric.
- `value` (`nvarchar(max)`) — the full text
- `txtlen` — length of the stored value
- `escaped` — the text contains HTML/XML escapes
- Joins: `subID`, `tagID`, `langID`, `dimID`, `coregID`
Size warning: `txt` is ~100 GB. The only index is the primary key on
`txtID`. Filtering by `subID` without the recommended `IX_txt_sub` index is a
full table scan.
### `cal` — Calculation linkbase
Parent/child tag arcs for the calculation relationships inside a filing.
Use this to reconstruct XBRL sums and reconcile parent values against their
components.
### `ren` — Rendering
Report-level metadata: `menucat` (`S` = statement, `N` = note, `D` =
disclosure), `shortname`, `longname`, plus the parent/child report
hierarchy (`parentreport`, `ultparentrpt`).
### `dim` — Dimensions (stripped)
**Important caveat:** `sec_fsnds.dim` carries only a `dimID` column. The raw
axis/member text (e.g.
`us-gaap:StatementBusinessSegmentsAxis / msft:ProductivityAndBusinessProcessesMember`)
is **not** stored in this database.
Rows in `num`/`txt` with `dimn > 0` are dimensional, and you can group them
by `dimID`, but you cannot decode which axis/member each `dimID` represents
without an additional mapping table. The original SEC `dim.tsv` file
contains that text and can be loaded separately if you need it.
### Lookup tables
All seven lookups share the same pattern: `xxxID` PK + one canonical text
column with a unique index. They exist purely to compress repeated strings
out of the hot tables.
| Lookup | Purpose |
|----------------|------------------------------------------------------------|
| `lkp_stmt` | Financial-statement category (`BS`, `IS`, `CF`, `EQ`, …) |
| `lkp_prole` | Preferred-role labels (`terseLabel`, `totalLabel`, …) |
| `lkp_uom` | Units of measure (`USD`, `shares`, `USD/shares`, `pure`) |
| `lkp_datatype` | XBRL datatypes (`monetaryItemType`, `textBlockItemType`) |
| `lkp_coreg` | Co-registrant name strings |
| `lkp_version` | Taxonomy namespace + version strings |
| `lkp_lang` | Language code |
### `fsnds` — Provenance
Tracks which raw SEC quarterly dump each row originated from. Every core
table carries an `fsndsID` column that FKs to this table, so you can
trace any fact back to its source TSV and load timestamp.
---
## Indexes present
Only the indexes below exist out of the box. Everything else you see in
the recommended-indexes script is additive.
| Table | Index | Keys |
|---------|-----------------------------------|-----------------------------------------------|
| `num` | `IX_num_tagID` | `tagID` INCLUDE(`subID,ddate,qtrs,dimID,coregID,value`) |
| `num` | `IX_num_sub_tag` | `subID, tagID` INCLUDE(`dimID,iprx,qtrs,ddate,uomID,value`) |
| `pre` | `IX_pre_sub_stmt` | `subID, stmtID` INCLUDE(`report,line,tagID,inpth,negating,plabel,proleID`) |
| `pre` | `IX_pre_label_remediation_bsicf` | Filtered on `inpth=0 AND stmtID IN (1,2,3)` |
| `sub` | `UQ_sub_adsh` | `adsh` (unique) |
| `sub` | `IX_sub_fy_cik` | `fy, cik` |
| `tag` | `UQ_tag` | `tag, versionID` (unique) |
| `tag` | `IX_tag_name` | `tag` INCLUDE(`tagID, abstract`) |
| `ren` | `IX_ren_sub_report` | `subID, report` INCLUDE(`menucat`) |
| `cal` | PK only |
| `dim` | PK only |
| `txt` | PK only *(see warning below)* |
| `tag_text` | PK only |
**Notable gaps** (see `sql/indexes/sec_fsnds_recommended_indexes.sql`):
- No filtered canonical index on `num` (`dimn=0 AND coregID IS NULL AND value IS NOT NULL AND iprx=0`).
- No index on `sub.sic` (peer queries scan all 820 K rows; still sub-second).
- No index on `sub.cik` or `(cik, form, filed)` (CIK-based filing searches scan).
- No index on `sub.name` (name-pattern searches scan).
- No index on `txt.subID`**required** before any practical use of
`usp_sec_get_text_facts`.
---
## Foreign keys
All core tables are fully FK'd back to `sub`, `tag`, the lookups, and
`fsnds`. The relationships follow the shapes above — see
`===FKS===` section of `sec_fsnds_schema_dump.out` or query
`sys.foreign_keys` directly if you need the exact list.
---
## Query idioms
Canonical consolidated numeric row:
```sql
WHERE n.dimn = 0 AND n.coregID IS NULL AND n.iprx = 0 AND n.value IS NOT NULL
```
Flow period classifier:
```sql
CASE
WHEN s.form = '10-K' THEN 4 -- annual flow
WHEN s.form = '10-Q' THEN 1 -- quarterly flow
ELSE NULL
END
```
Balance-sheet rows always use `qtrs = 0`. Income statement and cash flow
use `qtrs = 4` for 10-K and `qtrs = 1` for 10-Q. Q4 is never reported
directly — derive it from `10-K annual − (Q1 + Q2 + Q3 from 10-Qs)`, all
keyed on the same `sub.fy`.
---
## Where to go next
- `sql/indexes/sec_fsnds_recommended_indexes.sql` — opt-in indexes for the
common scan patterns.
- `sql/sps/` — 14 stored procedures covering single-filing retrieval,
time series, peer analysis, segment drill-down, and tag search.
- `sql/validation/usp_sec_validate_statement_integrity.sql` — bulk BS / CF /
IS sanity checker over arbitrary CIK + form slices.