Korean District Credit Risk Index

Overview

A monthly-updating district-level credit risk index for Korean residential real estate, built entirely from the 등기정보광장 (IROS) public court registry OpenAPI. The dataset covers 228 시군구 districts from January 2010 to the present across five registration types: 임차권 (court-ordered lease lien), 강제경매 (forced auction), 가압류 (provisional seizure), 전세권 (jeonse deposit right), and 근저당 (collateral mortgage). Each type serves a distinct role — 임차권 and 가압류 are tenant-initiated leading indicators; 강제경매 is the lagging foreclosure signal; 근저당 captures aggregate debt load. The headline empirical finding: a statistically consistent 17-month median lag between an 임차권 spike in a district and a subsequent 강제경매 spike in the same district (district-level R ≈ 0.93). This makes today's 임차권 trend a forward-looking predictor of foreclosure pressure 17 months out — and the core commercial proposition for 저축은행 and 상호금융 lenders evaluating district exposure.

Problem

The FSC's November 2025 supervisory revision assigns 110% weighting to 비수도권 저축은행 loan portfolios, creating a regulatory incentive for savings banks to expand into non-metropolitan districts. The data gap is one level up: which districts to expand into, and which carry deteriorating credit risk. Korean residential credit risk data is fragmented — KCB and NICE cover borrower scores, not district-level property registry trends. Transaction price data (MOLIT 실거래가) reflects completed deals, not the distress signals that precede defaults. No structured, time-series product exists that tracks the court registry leading indicators at the district level. That gap is the commercial opportunity.

Constraints

• IROS OpenAPI issues 1,000 calls per service key per day — a single key cannot complete a full backfill across 228 districts × 5 types × 197 months without hitting the daily limit, requiring 5 separate keys rotated across registration types
• DuckDB on Windows acquires an exclusive file lock — any process that holds the connection open blocks other reads; required explicit try/finally connection management throughout the pipeline
• The IROS API returns APIERROR-0003 (daily limit) silently mixed with valid responses — naively retrying on 0003 consumes the next day's quota; required a dedicated DailyLimitError class to skip remaining batches on limit hit
• 228 districts include 3 abolished administrative units (codes 405, 1511, 1610) that appear in the IROS codebook but return no data — required explicit handling to avoid phantom zero-rows poisoning the time series

Approach

The pipeline collects court registry data from the public IROS OpenAPI rather than transaction prices. Registration events are the earliest publicly-available signal of housing distress — a lease-lien is filed months before an auction is requested, which is filed months before it appears in the registry. By tracking all five registration types independently per district per month, the index captures the full distress lifecycle: debt accumulation (근저당), tenant exposure (임차권, 전세권), active dispute (가압류), and resolution (강제경매). A DuckDB fact table stores the time series; a SQLite state database logs every fetched API call for idempotent backfill. Monthly update runs via a single command (python -m pipeline update) and adds the prior two months. District-level composite scores weight the five types by signal reliability and apply rolling z-scores to normalise for baseline district activity.

Key Decisions

Tech Stack

• Python 3.11
• DuckDB (fact table + analytics)
• SQLite (idempotency state log)
• IROS 등기정보광장 OpenAPI (5 service keys)
• pandas, requests, python-dotenv

Result & Impact

• 228
  Districts covered
• 222,903
  Rows collected
• 16
  Years of history
• 5
  Registration types
• 74
  Tests passing
• ~0.93
  Cascade R²

The first publicly-derived, district-level Korean residential credit risk time series covering the full 임차권 → 강제경매 cascade. The 17-month median lag finding — reproducible from IROS public data by any analyst — gives 저축은행 and 상호금융 lenders a forward-looking signal that no commercially available product provides: today's lease-lien spike in a given 시군구 predicts, with strong statistical confidence, the foreclosure pressure that district will face 17 months from now.

Learnings

• Registry data is a structurally better signal for credit risk than transaction prices — it captures the distress pipeline, not just its outcome. This is the insight that makes the product differentiated rather than redundant with existing MOLIT-based tools.
• Idempotency at the API call level, not the batch level, is what makes a multi-month backfill survivable. A state log that records individual (district, type, month) triples costs almost nothing and eliminates the risk of corrupted partial batches.
• The 17-month lag is a district-level finding, not a property-level finding — it emerges from the aggregate series, not from tracking individual properties. Trying to apply it at the property level would be methodologically incorrect.
• Separated API keys per data dimension (one key per registration type) is a simple design choice that quadruples throughput under daily rate limits without any infrastructure complexity.