Overview

Half of Seoul's restaurants close within 6.1 years — but the naive "77% failure rate" everyone cites is statistically wrong.

Most public reporting on Seoul restaurant survival uses simple closure rates: "X% of restaurants fail." That number ignores restaurants still open (right-censoring) and treats a 1-year-old shop the same as a 10-year-old one. This project applies survival analysis — Kaplan-Meier curves, log-rank tests, and Cox proportional hazards regression — to 528,487 restaurant license records from all 25 Seoul districts, spanning 1980 to 2026. The corrected estimate (median survival of 6.1 years; 5-year survival of 55.9%) is more accurate than the naive framing, and the analysis surfaces three findings the naive rate hides: where you open matters more than what you cook, the COVID-era cohort is closing measurably faster, and restaurants that survive the first three years tend to persist.

Fig 1 · Overall Kaplan-Meier survival curve. Median = 73 months (6.1 years). The curve drops steeply in the first 3 years, then flattens — restaurants that survive the early period tend to persist.

Stack

The Problem

"What percentage of restaurants fail?" is the wrong question.

The framing in Korean small-business reporting has two specific statistical problems. First, it ignores right-censoring — 120,151 of the 528,487 restaurants in this dataset are still operating, and we don't yet know how long they'll last. Counting them as "succeeded" inflates the success rate; counting them as "failed" deflates it. Second, treating a 1-year-old restaurant the same as a 10-year-old one ignores exposure time. The right question isn't "how many failed?" — it's "at any given point in time, what's the probability a restaurant is still open?" That's a survival analysis question, not a closure-rate question.

Methodology

Five recorded decisions, each with its own Architecture Decision Record. The right tool first, defended choices second.

Every methodological choice in this project lives in a 03-DECISIONS.md file in the repo, written before the analysis was finalized. The goal: anyone reading the final notebook should be able to challenge the methodology by pointing at a specific decision, not by guessing.

Pipeline

2.27M raw nationwide records → 528,487 clean Seoul records, with six quality checks logged and verified.

The 658MB raw CSV from data.go.kr lands in cp949 encoding with 2,271,210 rows across all of South Korea. Filtering to Seoul drops it to 534,812; six quality checks then remove 6,325 problem rows (389 with license dates before 1970, 27 closures missing close dates, 2,156 missing district codes, plus duplicates and negative durations). What survives is 528,487 records — 408,336 closed, 120,151 still operating — covering all 25 Seoul districts.

Verification step before any analysis ran: all 25 districts present in the cleaned dataset, ~99% of closed stores had a close date (only 27 missing), zero negative durations, duration distribution looked plausible (median 59.5 months, p10 at 9.4 months). All of this is logged in the project's 03-JOURNAL.md as Phase-1 verification.

Findings

Three results that wouldn't appear in a naive closure-rate analysis: location dominates cuisine, COVID hit a measurable cohort, and survival is front-loaded.

1. Headline survival rates

The corrected (Kaplan-Meier) survival rates. The 5-year point estimate of 55.9% is roughly 22 percentage points higher than the naive "% closed" reporting would suggest.

2. Where you open matters more than what you cook

The median-survival gap across districts (5.3 years) is larger than the gap across cuisine types (4.6 years). City-center districts with tourism and office density (Jongno, Jung-gu) outperform residential suburbs (Yangcheon).

Fig 2 · KM curves by district. Jongno-gu (10.0y) at the top; Yangcheon-gu (4.7y) at the bottom. Log-rank p < 0.001.

Fig 3 · KM curves by business type. Chinese restaurants survive longest (9.3y); snack/street food the shortest (4.7y). Pattern tracks barriers to entry.

3. Cox PH: location effects are large and significant

Cox proportional hazards regression on a 100K sample, 14 covariates, all significant at p < 0.005. Jongno-gu is the most protective location (HR = 0.63, i.e. 37% lower closure risk than the baseline). Yangcheon-gu carries the highest risk (HR = 1.22).

Fig 4 · Cox PH hazard ratio forest plot (100K sample, 14 covariates). HR < 1 = lower risk; HR > 1 = higher risk. Concordance = 0.55 — significant but most variance still unexplained.

4. The COVID cohort — an unplanned discovery

Cohort analysis wasn't part of the original plan. It was added in Phase 3 to answer "are restaurants surviving longer or shorter than 5 years ago?" The result was sharper than expected.

Fig 5 · KM curves by 5-year opening cohort. The 2020–2024 cohort separates cleanly from the others — and more restaurants opened during COVID than in any prior 5-year window (~14K/yr vs. ~12K/yr earlier).

Limitations

What this dataset and methodology can't tell us — written explicitly so the analysis can be defended on its own terms.

• License date is a proxy for opening date. 인허가일자 is when the license was issued; actual opening can lag by weeks or months. Means the very-early-life curve is slightly conservative.
• No covariates beyond location and type. The license data has no revenue, no rent, no foot traffic, no nearby competition. That's why the Cox PH concordance is 0.55 — significant signal, but most variance is unexplained by what the data captures.
• Pre-2000 records may have survivorship bias. Older restaurants that closed long ago are less likely to have been digitized — early-period survival is slightly inflated.
• Competing risks not modeled. A restaurant closing and reopening under a new name shows up as two short-lived restaurants. The current model can't distinguish that from a true exit.

Lessons Learned

Picking the right statistical tool before running anything was worth more than any modeling afterward.

Tool choice is upstream of analysis. If I'd reported "% closed," I would have stated a ~77% failure rate — dramatic, cited, and statistically wrong, because it counts 120K still-being-observed restaurants as failures. The Kaplan-Meier estimate of 55.9% at five years is both more accurate and more useful for anyone actually deciding whether to open a restaurant. The lesson generalizes: picking the right tool for the right question is upstream of all the analytical work that follows.

Verify a dataset before designing around it. The data pivot from 골목상권 → 행정안전부 was avoidable. I'd planned the full project around 골목상권 (OA-15577) and didn't realize until I'd downloaded it that it was aggregate-level — quarterly counts of stores per commercial area, not individual store records. The 행정안전부 license data was the correct source from the start. The cost was a day of rework; the lesson is to inspect schemas before writing analysis plans.

What I'd do differently next time. Run the Cox PH on the full 528K dataset using chunked or out-of-core methods rather than falling back to a 100K random sample. The sample is statistically defensible, but it's a workaround I'd want to remove if I revisited this. I'd also test the proportional hazards assumption explicitly — location might matter more in the first 2 years than it does after — and consider modeling competing risks so a "closed-and-rebranded" restaurant doesn't get counted as two short-lived ones.

Links