Four in five survive.
And what that average hides Gyms (체력단련장업) · full licensing register, 27,908 records · survival analysis · public data

01What was measured

Korea's Ministry of the Interior publishes a local-government licensing dataset. It holds 27,908 facilities registered under 체력단련장업 — the licence category for gyms. Each row carries a licence date and, where applicable, a closure date. That is all a survival analysis needs: places still trading are censored at today, places that closed are events. Kaplan–Meier. As of 2026-09-03.

Why not simply divide closures by the total? A gym that opened two years ago has no five-year outcome yet. Drop it and only old gyms remain, so eras get mixed; count it as alive and the rate inflates. Kaplan–Meier counts it only as "alive through year two" and removes it from the denominator after that. The formula is one line — survival(t) = ∏ (1 − closures that day ÷ gyms alive that morning). It multiplies each day's "probability of surviving today", so a two-year-old gym contributes its first two years and the five-year figure is set only by gyms that completed five. The "closure rate in that year" in §04 is the same fraction taken year by year instead of multiplied — multiplied, you get cumulative survival; taken separately, you see when the risk rises.

Four in five survive five years, not one in five. "Survive" here means the registration is still alive — no closure or revocation has been filed — not that the gym is necessarily trading (§05 takes up that gap). It takes a little over fourteen years for half of them to close. That median moves with what you count as a closure: including administrative revocations gives 14.3 years, filed closures alone give 17.4. The five-year rate shifts by only a point under that choice; the median shifts by three years. So the number this piece leans on is the five-year one.

02I doubted my own number first

When a result overturns received wisdom fourfold, suspect the calculation before the wisdom. Three checks.

① Does a different method agree? Setting Kaplan–Meier aside, I simply counted, cohort by licence year, how many closed within five years. Across 22 cohorts (2000–2021) the median was 81.4%, range 76.6–85.7%. Stable across 22 years — not one period's artefact.

② Does another agency agree? I reconstructed the "currently trading" count at a past date from the raw register and compared it with the Ministry of Culture, Sports and Tourism's facility census. End-2023: 14,720 reconstructed against 14,773 published — 0.4% apart.

③ The one that mattered most — had old closures been purged? If the register discards long-closed records, only survivors remain and the rate is inflated wholesale. No modelling fixes that; only the data can settle it. Cumulative closure by licence year does.

Older cohorts sit higher and the decline is monotonic. The register holds twenty years of closures. Without this check the 82.6% would have been unusable.

03So what did the 18.4% count?

The figure the industry quotes comes from Statistics Korea's business demography series. It is not wrong. It counts something else.

The division mixes in sole traders and freelance instructors. A business started on a laptop and quietly abandoned does not trace the same curve as a site with a deposit, a fit-out, equipment and a filed licence. Both numbers are right. Attaching "gym" to the division-level figure is what is wrong. The figure returned by search matches the division value to the decimal — nobody re-measured a narrower population; a narrow label was fitted to a broad number. That is what makes it dangerous: it looks specific, so it is believed. To avoid repeating the error, this piece keeps the licence category name throughout.

04What the 82.6% average hides

Stopping here would make this a piece that corrected one number. The useful part is next — within the same licence category, who lives and who dies. I split it two ways.

Size. 20,947 records carry a floor area. Four bands.

Monotonic. The smallest and largest bands sit 13.5 points apart — the gap inside "gyms" is wider than the gap between gyms and the division. To rule out geography I re-ran it within Seoul alone: 79.9% against 92.7%, 12.9 points. Gyeonggi 11.1, Busan 26.3. Size separates them in every region.

Time. This was the opposite of what I expected. Received wisdom says new businesses are the risky ones. The share that close in a given year of life is lowest in year one.

Conditional closure rate (year n) = gyms that closed in that year ÷ gyms alive at the start of that year × 100. Not cumulative — it is the share of survivors that close in that year, so it shows how risk moves year by year.

Because recent openings may not have filed closures yet, I counted only the 12,137 gyms licensed 2000–2017, each observed for eight years or more.

Year-one risk doubles by year three and never comes back down. The risky gym is not the new one — it is the three-to-eight-year-old one.

The first explanation that comes to mind is the lease — if leases renew every two years, closures should pile up at 24 and 48 months. I recounted in six-month bins.

There is no spike at 24 or 48 months. It is not a contract calendar. The risk climbs twofold across the first two years and then stays at that height. That is as far as this register can go — the first two years are when the deposit and fit-out hold the door shut; once they are recovered, the gym is in a state where it can quit, and it stays in that state. Ageing equipment or a competitor next door are not in this data.

Overlay the two and it sharpens. Gyms under 100 m² close at 4.5% a year from year one. Gyms of 400 m² and over start at 0.9% and climb gradually to 3.3% by year six. Small ones are risky from the start; large ones become risky with age.

05What this number must not be used for

82.6% is close to an upper bound. A gym that stops trading without filing stays "open" in the register — an unfiled closure. I could not size that: of 16,853 currently open, 1,605 (9.6%) are over twenty years old, and since these are physical sites many genuinely last that long. Still, the direction is upward. A further 114 records (0.41%) are marked closed with no date and enter as censored; substituting gives 82.5%.

Pilates is not in here. It falls under separate categories. The label must be "gyms (체력단련장업)", not "fitness".

And there is an axis I threw away. The register has a "number of instructors" field, so I split on it, and gyms listing zero showed a strikingly low survival rate. A tidy story suggests itself at once — gyms without instructors fail. But opening those 352 records, 97% had no floor area either, and their median licence year was 1996. They were not gyms without instructors; they were old records nobody had filled in. That axis is not used.

Floor area is blank for 25% as well — 6,941 records. And the blanks are not random. The median licence year of a blank record is 2003 (2019 for filled ones), and their five-year survival is 73.7% against 86.0% for filled records — 12 points lower. Matched by era (opened 2000–2014) it is still 66.4% against 89.0%. The gyms that closed are the ones with the empty cells. So the size table above is drawn on "the filled-in 75% that survives better", and its absolute values run above the overall 82.6%. What to read there is not the level but the gap — and the gap survives splitting by era (10.5 points for 2000–2014 openings, 14.7 for 2015–2021).

06What changes because of it

For anyone selling to gyms — membership software, prepayment guarantees, equipment leasing — this data changes three things.

First: if four in five closed, you would be selling against failure risk; if four in five survive, that story does not hold. Nobody underwrites a market that collapses. Guarantees become a business precisely because most survive. So the risk lives in the variance, not the average — and the variance, as shown above, opens twofold on size alone. Charge a 30-pyeong gym and a 120-pyeong gym the same rate and one of them costs you money while the other walks. In numbers: the average five-year loss rate is 17.4%. Under 100 m² it is 22.5%, 1.3 times the average; 400 m² and over it is 9.0%, half. A rate set at the average under-collects by 30% on the small gyms and charges the large ones double. The ones paying double notice first — what stays is the side you under-charged.

Second: the timing of risk is backwards. Effort usually goes into screening new accounts, but the data says year one is the safest. Of the 3,839 gyms that closed within eight years, 7.9% closed in year one, 20.7% within two years, and 79.3% from year three on. Onboarding screening is aimed at 8% of the losses. Losses concentrate in existing accounts between years three and eight. A retention model should take tenure before it takes onboarding metrics, and the alarm should sound at the end of year two, not at signup.

Third: to use any of this, you clean first. Floor area is 25% blank, and reading a blank instructor count as zero manufactures a risk that does not exist. Public data is not usable on download; it becomes usable only after you have sorted what is missing from what is a blank pretending to be zero. Count without sorting and the conclusions change — keep the recent openings in and year-one closure reads 2.0% instead of 2.5%, safer than it is (their closures have not been filed yet); use "zero instructors" as it stands and you get a 56.5% risk that does not exist; treat the blank floor areas as random and you read the size table as if it were the whole population. Every table above came after that sorting.

Turned into actions. The thresholds are read straight off the tables above; redraw them from your own data.

If you are deciding something on an industry statistic, ask these four.

07What this piece is really saying

Nothing here is a new model. It counted something public that nobody had counted, and split what had been lumped into a single average. The register is downloadable by anyone; the dates and floor areas were in it from the start. But no one had drawn the curve, so an industry spent years holding another sector's number as its own — and never saw the twofold gap inside itself.

Most of our work looks like this. Far more often than data being missing, the data that exists is not in a state that can answer. Repeat-purchase rate cannot be computed until you decide what counts as one purchase; survival stays a single average until you decide what counts as one business and which blanks are pretending to be zero. Settling that comes before the analysis, and it is usually the part that is actually hard.

Getting data into a state where it can answer is what we do.

(Note — source: Ministry of the Interior local licensing data, category 체력단련장업, 27,908 records as of 2026-09-03. Kaplan–Meier; events are closures and revocations, trading and suspended are right-censored. Sensitivity with filed closures only: 83.7% at five years. Size bands use the 20,947 records (75%) with a floor area; 19 records listed at 10,000 m² or more were dropped as entry errors (one gym is recorded at 1,033,455 m²). Year-of-life closure rates use the 12,137 gyms licensed 2000–2017, observed eight years or more. Comparison: Statistics Korea business demography, arts/sports/recreation division, 2018 birth cohort. Zero internal data.)