DoDataThings/us-bank-transaction-categories

Update Notice

A newer version of this dataset is available: DoDataThings/us-bank-transaction-categories-v2

v2 adds [debit]/[credit] sign prefixes, 24,000 samples (up from 16,000), 500+ merchants, PayPal wrapper patterns across all categories, and a refined 16-category taxonomy (Housing split into Rent; Mortgage removed). See the v2 dataset card for details.

This v1 dataset remains available. If you don't need sign-aware data, v1 is still useful for basic transaction classification training.

US Bank Transaction Categories — Synthetic Dataset

16,000 synthetic transaction descriptions modeled after real US bank statement formats. Designed for training transaction classifiers that work on actual bank data — not the clean "Starbucks coffee" descriptions that most datasets use.

Why Synthetic?

Real bank transaction data is private. But the formats are universal — Chase, Apple Card, PayPal, Capital One all produce descriptions with predictable structures. This dataset captures those structures with randomized merchants, store numbers, addresses, and reference codes.

Format

CSV with two columns:

description,category
"AMAZON MKTPL*K8R2M5VN7","Shopping"
"Zelle payment to JAMES GARCIA, SEATTLE, WA 48291736502","Transfer"
"PAYPAL           INST XFER  NETFLIX         WEB ID: PAYPALSI42","Subscription"
"CHEVRON 0385291","Transportation"

Categories (16)

Category	Count	Examples
Restaurants	1,000	MCDONALD'S F28471, TST*SAKURA SUSHI - OAKLAND, SQ *POKE BOWL
Groceries	1,000	SAFEWAY #1197, WHOLE FOODS 3842 SEATTLE, SQ *TRADER JOE'S
Shopping	1,000	Amazon.com*K8R2M5VN7, TARGET T-2847, WALMART SUPERCENTER #3291
Transportation	1,000	CHEVRON 0385291, UBER *TRIP DENVER A8KM2Q, SHELL OIL 48291736
Entertainment	1,000	TICKETMASTER, AMC THEATRES #1234, STEAM
Utilities	1,000	PG&E, COMCAST XFINITY, SPECTRUM 48291
Subscription	1,000	PAYPAL INST XFER NETFLIX WEB ID:, GOOGLE *SPOTIFY, PP*APPLE.COM/BILL
Healthcare	1,000	CVS #4829, DR. SMITH FAMILY MEDICINE, WALGREENS PHARMACY
Insurance	1,000	FARMERS INS BILLING, STATE FARM PREMIUM, GEICO AUTO
Housing	1,000	MORTGAGE PAYMENT, AVALON RENT PMT, HOME DEPOT #0635
Travel	1,000	UNITED AIRLINES, MARRIOTT, AIRBNB, HERTZ
Education	1,000	COURSERA, BARNES & NOBLE #1234, UNIVERSITY OF SEATTLE TUITION
Personal Care	1,000	PLANET FITNESS, SEPHORA, SUPERCUTS #4829
Transfer	1,000	Zelle payment to NAME, CITY, ST, AUTOMATIC PAYMENT - THANK, CHASE CREDIT CRD AUTOPAY PPD ID:
Income	1,000	EMPLOYER PAYROLL PPD ID:, Zelle payment from NAME, INTEREST PAYMENT
Fees	1,000	OVERDRAFT FEE, ANNUAL MEMBERSHIP FEE, LATE PAYMENT FEE

Six Format Templates

The descriptions follow six real US bank statement formats:

1. ACH (Chase checking) — INSTITUTION PURPOSE DETAILS WEB/PPD ID: CODE
2. Merchant + ID (Chase credit) — MERCHANT #1234 or MERCHANT*ORDERID
3. Full address (Apple Card) — MERCHANT ADDRESS CITY ZIP STATE COUNTRY
4. PayPal prefix — PreApproved Payment Bill User Payment: MERCHANT
5. Action prefix (Capital One) — Withdrawal from DESCRIPTION
6. Simple — MERCHANT or MERCHANT.COM

Variation Dimensions

• Capitalization: ALL CAPS, Title Case, lowercase, mixed
• Spacing: normal, extra-padded (ACH style), compressed
• Store numbers: #1234, 1234, #01234, absent
• Order IDs: *ORDERID (Amazon style)
• POS prefixes: SQ * (Square), TST* (Toast)
• Addresses: full, partial, zip-smashed-into-city (Apple Card quirk)
• Cities: 36 US cities across multiple states

Usage

from datasets import load_dataset

ds = load_dataset("DoDataThings/us-bank-transaction-categories")
print(ds["train"][0])
# {'description': 'AMAZON MKTPL*K8R2M5VN7', 'category': 'Shopping'}

Design Decisions

• No "Business" category. Whether a transaction is a business expense depends on the account, not the description. An ANTHROPIC charge is "Subscription" regardless of which card it's on.
• Transfer vs Income. "Zelle payment to" = Transfer (money out). "Zelle payment from" = Income (money in). The preposition matters.
• Subscription vs Shopping. Amazon.com*ORDERID = Shopping (one-time purchase). AMAZON WEB SERVICES = Subscription (recurring SaaS). Format signals distinguish them.
• Balanced classes. 1,000 samples per category prevents the classifier from defaulting to the most common class.

Trained Model

A DistilBERT model fine-tuned on this dataset is available at DoDataThings/distilbert-us-transaction-classifier — 99.75% validation accuracy, 90% on real unseen bank descriptions.

Generator

The synthetic data generator script is open source. You can extend the merchant pools, add new format templates, or increase the sample count:

node scripts/generate-training-data.js --count 1000

Available at github.com/wnstnb/foliome.

License

MIT