A tidier and open-data version of CDB90 (CAA Database of Battles, Version 1990). This used to be called CDB13, which I thought was an amusing name, but seemed to make people think I added to the CDB90 data.
The description of CDB90 from Helmbold (1993):
A database of over 600 battles that were fought between 1600AD and 1973AD. Descriptive data include battle name, date, and location; the strengths and losses on each side; identification of the victor; temporal duration of the battle; and selected environmental and tactical environment descriptors (such as type of fortifications, type of tactical scheme, weather conditions, width of front, etc.)
The DTIC record for the database (AD1038724, April 30, 1991, approved for public release) is in src-data/docs, with an OCR text made with LlamaParse, and the OCR text of the six HERO volumes (CAA-SR-84-6, which describe the first 600 battles) is in src-data/docs/hero (uv run dvc unfreeze ocr && uv run dvc repro ocr, with LLAMA_CLOUD_API_KEY set).
The original documentation for the CDB90 dataset is in src-data/M000121/README.TXT.
The directory src-data/M000121 contains the original data from CDB90 as obtained from the NTIS.
The original datafiles are the WKS files.
However, since this is an archaic format, these were converted to csv files using LibreOffice (v. 4.4.3.2; BuildId: 88805f81e9fe61362df02b9941de8e38a9b5fd16; Locale: en_).
The directory data contains the revised data (see Formats).
No major substantive changes were made to the data, nor were new battles added.
A brief summary of the revisions are:
- Minor changes to several values so that logical relationships hold. These are all the edits made in the revision CDB91 as described in Helmbold (1995) DTIC reort ADA298124, p. 2-4. The remainder are a few other logical inconsistencies in the original data: e.g. total tanks must equal light battle tanks plus main battle tanks; tanks are 0, not missing prior to WWI.
- Reformatting the data from one of the messier spreadsheets that I have ever seen to tidy data. This means that there are separate tables for battles, combatants, activity periods, etc.
- Additional variables linking the battles to other datasources such as the Correlates of War
- Different categorizations of wars, both from the original HERO dataset and my own categorization which is mostly consistent with the COW wars after 1815, and Wikipedia before 1815.
The flat CSV tables are described by a Data Package (datapackage.json).
There are three ways to get the data, all in version 10.0.0 and later (from the GitHub release and Zenodo).
| Format | Files | For |
|---|---|---|
| Flat CSV and Parquet | data/flat/battles.{csv,parquet}, data/flat/sides.{csv,parquet} |
Analysis: one row per battle, or per battle and side, with the tables already joined. |
| SQLite | cdb90.sqlite3 (release asset) |
The relational tables: battles, belligerents, commanders, active periods, weather, terrain, front widths, battle links, commander links, coordinates, notes on source values and the code tables. |
| Relational CSV | data/relational/*.csv |
The inputs of the SQLite database, one csv per table. |
datapackage.json is a Frictionless Data Package for the flat CSV tables.
battles: one row per battle, with its dates and duration, coordinates, most specific Wikipedia link and per-side totals. The coded variables (wina,cea,surpa, ...) are from the attacker's point of view.sides: one row per battle and side, with forces, commanders and the outcome from that side's point of view (1 win, 0 draw, -1 loss).
All 660 battles are linked to one or more Wikidata items and Wikipedia articles (a battle's own article where there is one, otherwise the smallest campaign or war article that covers it) (battle_links in SQLite, wikidata_id and wikipedia_title in the flat battles table), and all 660 battles have coordinates from Wikidata (coordinate_precision says whether the point is the battle, the place where it took place, a town or district beside the action, or only its campaign or region). Two points are best guesses because the engagement could not be identified: Chan River (isqno 632) and Bayonette Line (636).
1,158 of the 1,263 named commanders are linked to a Wikidata item that identifies the person CDB90 lists (commander_links). A link says who the name refers to, not that the person really held the command. The other 105 have no link because they could not be identified or have no item, mostly colonels and majors.
Values in the source data are kept as they are. Where another source shows one to be wrong, data_notes records the value, the problem and the source, including wrong Wikipedia links in commanders.uri.
The narratives of the HERO volumes (the short account of each battle in CAA-SR-84-6) and of the 1988 report New Engagement Data for the Breakpoints Data Base (battles 602 to 625) are parsed from the OCR text into data/flat/narratives and the battle_narratives table: for 610 battles, the narrative, its significance and its sources, with each source code of the volumes replaced by the full citation (the 1988 report lists its sources by operation) and a document column that says which report each narrative is from. The OCR is not perfect, so check a name or a number against the PDF (src-data/docs/README.md).
Version 10.0.0 moved the relational csv tables from data/ to data/relational/.
See CITATION.cff. The data (not the code) is archived on Zenodo:
the SQLite database, the flat CSV and Parquet tables, the relational CSV tables as one zip, the data
package, the codebook and the original documentation. The DOI 10.5281/zenodo.23928303 always points to
the latest version; version 10.5.0 is 10.5281/zenodo.23931975 (10.0.0 is 10.5281/zenodo.23928305).
Arnold, Jeffrey B. (2026). CDB90: A tidy, open-data version of the CAA Database of Battles (1990) (Version 10.5.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23931975
To prepare a deposit, uv run python -m scripts.zenodo bundle collects the files in dist/zenodo/,
and uv run python -m scripts.zenodo deposit uploads them as a draft (set ZENODO_TOKEN; --update ID
replaces the files of an existing draft; --publish publishes, which cannot be undone). A new version of a
published record is made with "New version" on Zenodo, which the script does not do yet.
- CAA Study Report CAA-SR-84-6, "Analysis of Factors That Have Influenced Outcomes of Battles and Wars: A Data Base of Battles and Engagements," September 1984: AD-B086-797L, AD-B087-718L, AD-B087-719L, AD-B087-720L, AD-B087-721L, AD-B087-722L
- HERO Report Number 129, "Combat History Analysis Study Effort (CHASE) Data Enhancement Study (CDES)," 31 January 1986, AD-A175-712, AD-A175-713, AD-A175-714, AD-A175-715, AD-A175-716
- Data Base Error Correction (DBEC)," 23 January 1987. It was prepared for CAA under Purchase Order Number MDA903-86-M-8560 and is available from DTIC under accession number AD-A176-750.
- FW Management Associates, Inc. Report "Independent Review/Reassessment of Anomalous Data (IR/RAD)," 22 June 1987, in four volumes. It was prepared for CAA under Contract Number MDA903-86-C-0396. It is available from DTIC under the following accession numbers: AD-???-??? (Volume I), AD-195-726 (Volume II), AD-???-??? (Volume III), and AD-???-??? (Volume IV)
- Data Memory Systems Incorporated report, "New Engagement Data for the Breakpoints Data Base," prepared for the US Army Concepts Analysis Agency under Contract No. MDA903-87-C-0807, 30 September 1988.
- Robert Helmbold. 1993. "Personnel Attrition Rates in Historical Land Combat Operations: A Catalog of Attrition and Casualty Data Bases on Diskettes Usable with Personal Computers." DTIC ADA279069.
- Robert Helmbold. 1995. "Personnel Attrition Rates in Historical Land Combat Operations. Some Empirical Relations among Force Sizes, Battle Durations" DTIC AD-A268 787.
- Code is MIT, unless otherwise noted.
load-sqlite.pyis vendored from okfn/dptools; upstream gave no license. - The derived data in
data/anddatapackage.jsonis CC0 1.0. - Battle links and coordinates come from Wikidata, which is CC0.
src-data/local/battle_links.csvis curated by hand and with Wikipedia's search. - The original CDB90 data in
src-data/M000121is in the public domain.