README.md

metadata

license: cc-by-4.0
pretty_name: RecBench
tags:
  - finance
  - reconciliation
  - entity-matching
  - synthetic
size_categories:
  - 100M<n<1B

RecBench — an open benchmark for financial transaction reconciliation

RecBench is a synthetic, fully-reproducible benchmark for the core problem of financial operations: matching a company's internal ledger against processor settlement reports and bank statements, at the scale large fintechs operate (up to 100M transactions/year), with exact ground truth and a standard evaluation.

Built by End Close. Generator code Apache-2.0; datasets CC BY 4.0.

Why

No open reconciliation benchmark exists. Vendors publish only marketing match-rate claims. RecBench provides what the adjacent entity-matching field has had for a decade (Magellan, Leipzig, WDC): public data, exact labels, canonical metrics, published generator.

The task

Three-way reconciliation over three record collections:

• ledger/ (side A) — the company's internal view: charges, refunds, disputes, across 6 currencies and 3 processors
• settlement_lines/ + payouts/ (side B) — processor reports: gross/fee/net arithmetic, FX conversion, timezone sales-day batching
• bank_statements/ (side C) — deposits net of everything, degraded references, noise lines

A submission must (1) pair ledger transactions with settlement lines, (2) pair payouts with bank deposits, (3) un-bundle deposits whose settlement report is missing (subset-sum matching), and (4) flag the records that cannot be matched. Difficulty spans exact matches through FX/fee arithmetic, date-boundary shifts, partial captures, duplicated lines, missing references, and provably-ambiguous cases — every case labeled, so results break down by difficulty type. See docs/DIFFICULTY.md.

Tiers

tier	ledger rows	total rows (all tables)	size (zstd parquet)	generation time*	seed
small	1.0M	~4.3M	67 MB	~20 s	1042
medium	10.0M	~43M	644 MB	~2 min	1043
large	100.1M	~430M	6.0 GB	~19 min	1044

* single core, Apple Silicon. All tiers simulate one year plus a 14-day settlement tail; the large tier has ~3.6M payouts (53k of them subset-sum unbundling cases). Any tier regenerates byte-identically from its seed:

python3 -m venv .venv && source .venv/bin/activate     # Python >= 3.11
pip install -e .
recbench generate --tier small --out data/small        # canonical seed
recbench generate --tier small --seed 99 --out data/x  # fresh private set

Evaluation

python eval/evaluate.py --dataset data/small --submission my_submission/

Headline metric: auto-match rate at ≥ 99.9% pair precision (industry convention — a false match is far costlier than an exception), plus precision/recall/F1 per difficulty slice, payout-matching and deposit-unbundling scores, and exception-detection accuracy. Submission format: eval/submission_format.md. Reference baselines (eval/baselines/): a perfect oracle and a naive exact-rules matcher that motivates the benchmark:

metric	naive baseline
ledger↔settlement F1	0.996
auto-match rate @ P≥99.9%	0.000 — duplicate lines poison its top confidence tier
missing-reference recall	0.644
deposit unbundling (group-exact)	0.000
exception recall	0.000

The last few percent are the benchmark.

Contamination-resistant evaluation

Ground truth for canonical tiers is public for reproducibility. To evaluate systems that may have trained on it, mint a fresh set from a private seed — same difficulty distributions, entirely different records. Please keep the truth_*/labels_* files out of crawlable training corpora.

Repository layout

recbench/          seeded deterministic generator (config in recbench/config/)
eval/              evaluator, submission format, reference baselines
docs/              SCHEMA, DIFFICULTY, EVALUATION, METHODOLOGY, DATASHEET
scripts/           canonical release build + Hugging Face upload
tests/             determinism, accounting-invariant, and evaluator tests

Citing

RecBench: an open benchmark for financial transaction reconciliation. End Close, 2026. https://endclose.com

Versioning

Dataset version = generator version (currently 0.1.0, pre-release). Results are comparable only within a version; every dataset's manifest.json records version, seed, config hash, and per-file sha256.