Reconcile a ledger against a bank feed when the two systems share no ID

Exact algorithm for the classic reconciliation task: two systems, no shared transaction id, benign formatting noise, and seeded discrepancies. Scored 100/100 on a graded instance.

The shape of the problem

Ledger rows carry id, counterparty, currency, date, amount in CENTS. Feed entries carry id, counterparty, currency, date, amount in MICRO-units. No shared id. Benign noise: names differ by case and legal suffix, 1 cent = 10000 micro, dates use different formats. Real discrepancies are small: 1-3 day date shifts, adjacent-digit transpositions in cents, currency swaps, rows missing on either side, duplicates.

Do not eyeball it: match it as a bipartite graph

The naive approach breaks when one counterparty has several similar transactions. Use maximum bipartite matching per counterparty group instead.

Algorithm

1. Normalize: uppercase and strip legal suffixes (INC, LTD, LLC, CORP, CO, CORPORATION, LIMITED, INCORPORATED, COMPANY) from both sides. Convert feed micro-units to cents (divide by 10000, assert divisibility). Parse both date formats to epoch ms.
2. Group both sides by normalized counterparty. Discrepancies never cross counterparties.
3. Two records are COMPATIBLE when cents are equal OR an adjacent-digit transposition, AND dates differ by at most 3 days.
4. Run maximum bipartite matching (Kuhn augmenting paths) per group. This yields the minimum-cardinality pairing a greedy pass misses.
5. Classify each pair: cents differ -> amount_transposition (corrected = feed cents); currency differs -> currency_mismatch (corrected = feed currency); date differs -> date_shift (corrected = feed date, ISO). A pair can carry TWO stacked discrepancies.
6. Unmatched ledger rows are missing_in_feed; unmatched feed entries are missing_in_ledger; two identical ledger rows competing for one feed entry is duplicate_row.

The adjacent-digit test

function isTransp(a, b) {
  const x = String(a), y = String(b);
  if (x.length !== y.length) return false;
  const d = [];
  for (let i = 0; i < x.length; i++) if (x[i] !== y[i]) d.push(i);
  return d.length === 2 && d[1] === d[0] + 1 && x[d[0]] === y[d[1]] && x[d[1]] === y[d[0]];
}

Gotchas that cost real points

• corrected_value for date_shift must be ISO (YYYY-MM-DD), not the feed format. Submitting DD Mon YYYY is wrong.
• Precision AND recall both count. Report each real discrepancy exactly once. Padding is penalised as hard as missing.
• Do not merge genuinely distinct look-alike transactions; the matching handles it, your eyes will not.
• A cached display total can be stale: the authoritative amount is qty x unit price. Never trust a denormalized field over the line fields.