VaidhyaMegha/banking-kg

Banking Knowledge Graph — a bank and its market

75,179 nodes. 241,222 edges. Washington DC's 2023 lending market from four public regulatory filings, and one fictional bank's 5,000 customers, joined where they honestly can be.

Built with Samyama Graph. Source and loader: samyama-ai/banking-kg.

Two layers: one real, one synthetic

Layer	Real or synthetic	Nodes	Edges
Market — DC 2023 (HMDA, GLEIF, FDIC, SBA)	real, public	31,128	83,640
Bank — Banking-KG Bank, 5,000 customers, as of 2026-08-31	synthetic	44,044	152,795
Campaign audiences (computed)	synthetic	6	3,762
Bridge — the bank as one more market lender; its mortgages in HMDA's product taxonomy	synthetic bank node	1	1,025
Total		75,179	241,222

Every synthetic node carries synthetic: true; no real node does. Banking-KG Bank, its customers and its competitors are fictional, and are never matched to a real lender or person.

Why a graph

A bank's own book and its regulatory filings live in different tables, from different regulators, keyed differently. As a graph, "where did this number come from?" is a path:

// One mortgage traced to its application, lender, legal entity, Call Reports, tract and buyer
MATCH (a:Application)-[:ORIGINATED_AS]->(n:Loan)-[:SOLD_TO]->(p:Purchaser)
MATCH (n)-[:ORIGINATED_BY]->(l:Lender)-[:REGISTERED_AS]->(e:LegalEntity)
MATCH (a)-[:IN_TRACT]->(t:Tract)
MATCH (l)-[:FILED]->(f:Filing) WHERE f.period IN ['2022Q4', '2023Q4']
RETURN a, n, p, l, e, t, f LIMIT 25

// The bank inside its market: our 2023 mortgages and the market's, at the same LoanProduct
MATCH (c:CustomerLoan)-[:CLASSIFIED_AS]->(p:LoanProduct)<-[:FOR_PRODUCT]-(a:Application)
RETURN p.name, count(DISTINCT c) AS ours, count(DISTINCT a) AS market

// A family starts to leave: one member closed accounts, the others are still active
MATCH (gone:Customer {status: 'CLOSED'})-[:MEMBER_OF]->(h:Household)<-[:MEMBER_OF]-(c:Customer {status: 'ACTIVE'})
RETURN h.id, gone.name, collect(c.name) AS still_with_us LIMIT 10

Files

Nodes carry an id; edges reference those ids as src and tgt. banking-kg.sgsnap is the complete graph, including 384-dimension embeddings on 930 nodes (all-minilm) that the CSVs leave out.

nodes/

File	Rows
card.csv	20,698
application.csv	17,474
account.csv	11,525
loan.csv	8,616
customer.csv	5,000
household.csv	2,677
customerloan.csv	1,807
collateral.csv	1,517
sbaloan.csv	966
filing.csv	895
business.csv	858
address.csv	807
creditline.csv	707
lender.csv	539
legalentity.csv	478
industry.csv	221
tract.csv	201
franchise.csv	44
merchantcategory.csv	35
product.csv	32
counterparty.csv	29
businessunit.csv	17
denialreason.csv	9
purchaser.csv	9
audience.csv	6
loanproduct.csv	6
developmentcompany.csv	4
county.csv	1
msa.csv	1

edges/

File	Rows	Connects
spends_at.csv	81,269	Account → MerchantCategory
has_card.csv	20,698	Account → Card
filed_with.csv	17,474	Application → Lender
for_product.csv	17,474	Application → LoanProduct
in_tract.csv	17,346	Application → Tract
owns.csv	14,670	Customer → Account
of_product.csv	11,525	Account → Product
originated_as.csv	8,616	Application → Loan
originated_by.csv	8,616	Loan → Lender
owes.csv	5,777	Customer → Counterparty
sold_to.csv	5,114	Loan → Purchaser
member_of.csv	4,300	Customer → Household
in_audience.csv	3,762	Customer → Audience
denied_for.csv	3,718	Application → DenialReason
holds_card.csv	3,603	Customer → Card
paid_by.csv	2,803	Account → Counterparty
sends_to.csv	2,178	Account → Counterparty
borrows.csv	1,807	Customer → CustomerLoan
has_loan.csv	1,807	Account → CustomerLoan
secured_by.csv	1,517	CustomerLoan → Collateral
classified_as.csv	993	CustomerLoan → LoanProduct
borrowed.csv	966	Business → SBALoan
made_by.csv	966	SBALoan → DevelopmentCompany, SBALoan → Lender
filed.csv	895	Lender → Filing
in_industry.csv	866	Business → Industry
located_at.csv	844	Business → Address
has_credit_line.csv	707	Account → CreditLine
registered_as.csv	478	Lender → LegalEntity
in_county.csv	201	Tract → County
receives_from.csv	134	Account → Counterparty
under_franchise.csv	65	SBALoan → Franchise
offered_by.csv	32	Product → Lender
in_msa.csv	1	County → MSA

Sources and licences

Source	Publisher	Terms
HMDA Loan/Application Register 2023 (DC)	CFPB / FFIEC	U.S. government public data
FDIC BankFind institutions and financials	FDIC	U.S. government public data
SBA 7(a) & 504 FOIA (FY2020+, FY2010+)	SBA	U.S. government public data
LEI records	GLEIF	CC0 1.0
Bank layer	Samyama synthetic generator (seeds 42 and 7)	synthetic; no real person or institution

The sources keep their own terms. The compiled graph, schema and code are © 2026 Samyama.ai, under the terms in LICENSE.

Usage

from datasets import load_dataset
lenders = load_dataset("VaidhyaMegha/banking-kg", "lender", revision="v1.0")

docker run --rm -p 8080:8080 -p 6379:6379 public.ecr.aws/f9f6l5u4/samyama-graph:1.1.0
curl -X POST localhost:8080/api/snapshot/import -F "file=@banking-kg.sgsnap"

Vector search and natural-language query also need the tenant's embed_config and nlq_config, which are not part of a snapshot; see the source repository.

Verification

• Totals reconcile against the snapshot header: 75,179 nodes, 241,222 edges, 29 labels, 33 edge types.
• banking-kg.sgsnap sha256 e70f56e210f236eba660fab2e4b63615aa814f8ef96cbcb1319537fbb11b72fb, the value recorded in the source repository's data/SNAPSHOT.md (exported 2026-10-07, engine 1.7.1).
• 0 dangling edges. 17 orphan nodes: the 17 BusinessUnit nodes, which the customer data request gives no way to join to anything.
• Market: HMDA 17,474 records re-verified against the live endpoint (2026-10-06); triangle count 8,616 equals the number of loans (every loan closes application → loan → lender).
• Bank: the generator's validator passes (0 schema, key or code problems; 14 of 14 cross-table checks).

Limitations

• No effect sizes from the bank layer: its patterns were authored and planted, not learned.
• HMDA shows lending decisions, not loan performance, and the public file has no credit scores: do not draw credit or fair-lending conclusions about named lenders from it.
• LEI → FDIC certificate is a legal-name match (unique matches only). HMDA purchaser is a category, not an institution. Tract adjacency is not built.
• The bank's customers live in VA, MD and NC; the market layer is DC.

Citation

Banking Knowledge Graph, v1.0 (75,179 nodes, 241,222 edges). Market layer: DC 2023 public filings
(HMDA, FDIC, SBA, GLEIF). Bank layer: synthetic. Built with Samyama Graph.
Source: https://github.com/samyama-ai/banking-kg