README.md

metadata

license: mit
tags:
  - text-classification
  - finance
  - transactions
  - english
  - synthetic
task_categories:
  - text-classification
language:
  - en
size_categories:
  - 10K<n<100K

US Bank Transaction Categories v2 — Synthetic Dataset

68,000 sign-prefixed transaction descriptions across 17 spending categories, modeled after real US bank statement formats. Designed for training classifiers that work on actual bank data — not the clean "Starbucks coffee" descriptions that most datasets use.

Successor to v1.

Why This Dataset

Real bank transaction data is private. But the formats are universal — Chase, Apple Card, PayPal, Capital One, Mercury all produce descriptions with predictable structures. This dataset captures those structures with 500+ real merchant names, randomized store numbers, addresses, and reference codes.

Why Sign Prefixes

Bank transaction descriptions are ambiguous without direction. "VENMO CASHOUT" as a credit is income — as a debit it's a transfer. [debit]/[credit] encodes the cardholder's perspective so the model can disambiguate.

Sign distributions per category reflect real-world patterns:

• Spending categories: 95%+ debit
• Income: 97% credit
• Transfer: 50/50 (bidirectional by nature)

Format

CSV with two columns:

description,category
"[debit] AMAZON MKTPL*K8R2M5VN7","Shopping"
"[credit] ACME CORP       PAYROLL                    PPD ID: 123456789","Income"
"[debit] PreApproved Payment Bill User Payment: Netflix","Subscription"
"[debit] SCHWAB BROKERAGE MONEYLINK PPD ID: 5551234567","Transfer"
"[credit] AUTOMATIC PAYMENT - THANK","Transfer"
"[debit] PP*SAFEWAY","Groceries"
"[debit] TST*GOLDEN DRAGON - OAKLAND","Restaurants"

Categories (17)

Category	Count	Examples
Restaurants	4,000	TST*SAKURA SUSHI, DAVESHOTCHICKEN, CLV*BURGER JOINT, PP*CHIPOTLE
Groceries	4,000	SAFEWAY #1197, WHOLEFDS FRE #10467, PYPL*TRADER JOE'S, 365 MARKET
Shopping	4,000	Amazon.com*K8R2M5VN7, TIKTOK SHOP, Express Checkout Payment: TARGET, TOTAL WINE
Transportation	4,000	CHEVRON 0385291, CHARGEPOINT *STATION 1234, COSTCO GAS #1061, CA DMV FEE
Entertainment	4,000	VALVE STEAM PURCHASE, DRAFTKINGS SPORTSBOOK, PreApproved Payment: Valve Corp.
Utilities	4,000	PG&E, PGANDE WEB ONLINE, REPUBLIC SERVICES TRASH, SUNRUN SOLAR
Subscription	4,000	CURSOR USAGE, X CORP. PAID FEATURES, PYPL*NETFLIX, SALESFORCE; Billing
Healthcare	4,000	TELADOC TELEHEALTH, KAISER PERMANENTE, PP*ASPEN DENTAL, BETTERHELP
Insurance	4,000	FARMERS INS BILLING, HOMESERVE USA, PreApproved Payment: STATE FARM
Mortgage	4,000	ROCKET MORTGAGE, PATELCO CU MORTGAGE, SOFI MORTGAGE PAYMENT, Principal Pmt
Rent	4,000	EQUITY RESIDENTIAL RENT PAYMENT, GREYSTAR LEASE PAYMENT
Travel	4,000	ROYAL CARIBBEAN, TSA PRECHECK, SWA INFLIGHT WIFI, MARRIOTT
Education	4,000	COURSERA, Express Checkout Payment: Scholastic Inc, BRILLIANT.ORG
Personal Care	4,000	SPORT CLIPS, SEPHORA, 360 FITNESS LLC, TST*ISLAND VINTAGE SHAVE
Transfer	4,000	COINBASE ACH TRANSFER, WIRE TRANSFER TO NAME, ATM DEPOSIT, DDA TO DDA
Income	4,000	PAYROLL PPD ID:, SSA TREAS 310 FED SAL, DOORDASH DASHERPAY
Fees	4,000	OVERDRAFT FEE, ATM SURCHARGE, PAPER STATEMENT FEE

Eight Bank Statement Formats

Every spending category produces descriptions in all major US bank formats:

Format	Structure	Banks
Chase ACH	INSTITUTION PURPOSE PPD/WEB ID: CODE	Chase checking
Chase merchant	MERCHANT #STORE or MERCHANT*ORDERID	Chase credit cards
Apple Card	MERCHANT ADDRESS CITY ZIP STATE COUNTRY	Apple Card
PayPal	PreApproved Payment: MERCHANT, PP*, PYPL*, PAYPAL *, INST XFER	PayPal (as card issuer)
Capital One	Withdrawal from MERCHANT, Preauthorized Deposit from MERCHANT	Capital One
Mercury	MERCHANT; Description or just MERCHANT	Mercury, neobanks
POS	SQ *MERCHANT, TST*MERCHANT, CLV*MERCHANT	Square, Toast, Clover
Simple	MERCHANT, MERCHANT.COM, MERCHANT INC.	Various

PayPal as a Bank Format

PayPal isn't just a payment wrapper — it's a card issuer. People use PayPal credit/debit cards at any merchant. This dataset treats PayPal formats as first-class bank statement structures:

• PreApproved Payment Bill User Payment: STARBUCKS → Restaurants
• PP*SAFEWAY → Groceries
• PYPL*NETFLIX → Subscription
• Express Checkout Payment: TARGET → Shopping
• PAYPAL INST XFER MEDIUM.COM WEB ID: PAYPALSI77 → Subscription

PayPal-formatted descriptions appear in all spending categories at realistic rates (5-15% of samples).

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 *, TST*, CLV*
• PayPal prefixes: PP*, PYPL*, PAYPAL *, PreApproved Payment, Express Checkout
• Addresses: full, partial, zip-smashed-into-city (Apple Card quirk)
• Compressed names: DAVESHOTCHICKEN, CHICKFILA, WHOLEFDS FRE, TRADERJOES
• Cities: 36 US cities across multiple states
• Sign prefixes: [debit]/[credit] with category-appropriate distributions

Design Decisions

• No "Housing" category. Split into Mortgage (model-classified) and Rent (model-classified). Home maintenance → Shopping.
• No "Business" category. Whether a transaction is a business expense depends on the account, not the description.
• Transfer vs Income uses sign. [credit] VENMO CASHOUT = Income. [debit] VENMO PAYMENT TO = Transfer. The sign prefix is the primary disambiguator.
• Subscription vs Shopping uses format cues. Amazon.com*ORDERID = Shopping. AMAZON WEB SERVICES = Subscription. X CORP. PAID FEATURES = Subscription. TIKTOK SHOP = Shopping.
• 500+ real merchants across restaurants (including culturally diverse cuisines), groceries, SaaS/AI tools, fintech platforms, crypto brokerages, EV charging, gambling/sportsbooks, and more.
• Balanced classes. 4,000 samples per category prevents the classifier from defaulting to the most common class.

Usage

from datasets import load_dataset

ds = load_dataset("DoDataThings/us-bank-transaction-categories-v2")
print(ds["train"][0])
# {'description': '[debit] AMAZON MKTPL*K8R2M5VN7', 'category': 'Shopping'}

Trained Model

A DistilBERT model fine-tuned on this dataset is available at DoDataThings/distilbert-us-transaction-classifier-v2 — 99.9% validation accuracy, 96% of real-world classifications at 0.90+ confidence.

Generator

The synthetic data generator is open source:

node scripts/generate-training-data.js --count 4000  # 4,000 per category

Available at github.com/wnstnb/foliome.

License

MIT