Ma Dong-seok and Park Ji-hwan: the 1.9% who repeat

Ma Dong-seok and Park Ji-hwan have appeared in four of the same Netflix Top 10 titles. Jo Woo-jin and Kim Eui-sung have too. In this data they are the only two pairs who have done it four times — and they are the exception, not the rule.

Take every Korean film and series that has reached a Netflix Top 10 and has a cast recorded in Wikidata — 666 titles, 1,391 people, 3,579 credited roles. Draw a line between two people every time they appear in the same production.

You get 14,144 distinct pairs. Of those, 268 have worked together more than once. That is 1.9%.

The two pairs at four are Jo Woo-jin with Kim Eui-sung, and Ma Dong-seok with Park Ji-hwan. Across 635 productions, nobody in this data has shared a credit with the same person five times.

Chance would give you almost the same number

1.9% sounds low, but low compared to what? So we built the comparison: keep every title’s cast size exactly as it is, pour all 3,579 roles into one pile, and deal them back out at random.

A random Korean film industry produces about 214 repeat pairs. The real one produces 268.

Thirty shuffles gave a median of 214 and a range of 194 to 253 — so the observed 268 sits above the middle of the random distribution but inside its range, and one shuffle in thirty beat it outright. Observed is 1.3 times chance. That is a real effect and a small one.

What that rules out

It rules out the repertory company as a description of the whole industry. If Korean production ran on a few tight circles who cast each other repeatedly, this ratio would not be 1.2. It would be several times chance, and the four-time pairs would be commonplace rather than two of them.

It does not rule out that such circles exist. Ma Dong-seok and Park Ji-hwan have made four of these titles together, which is not chance for those two people. The finding is about the aggregate: the circles are too few and too small to bend the whole distribution.

What could hide a repertory habit from us

Three things, and all of them push in the same direction — they make the real figure higher than 1.8%, not lower.

Wikidata’s cast lists are short. The median title here lists four people. A real production has dozens of credited actors, and the ones Wikidata records are the well-known ones. Two character actors who work together constantly can share ten films and never produce a single pair in our data because neither is listed.

We only see titles that charted. A director who casts the same four people in everything will show up here only for the productions that reached a Netflix Top 10. The rest of that body of work is outside the panel.

And we cannot see behind the camera. P161 is cast. If Korean production has a repertory habit, the most likely place for it is a director returning to the same crew, and none of that is in this data.

So read 1.8% as a floor on repetition and a ceiling on how confidently we can talk about it. What we can say is narrower: in the part of the industry that is visible in charting titles and recorded casts, actors pair up close to randomly.

Why we ran the shuffle at all

Because “only 1.8% repeat” is the kind of number that reads as a finding and is not one on its own. A sparse network of 1,355 people spread over 635 small casts will produce very few repeats no matter how it is organised. Without the null model we would have published a fact about arithmetic and called it a fact about Korean casting.

The shuffle is what turns it into a claim: not “repeats are rare” but “repeats are barely more common than they would be if nobody chose anybody.”

---

If pairings are close to random, what holds the network of Korean productions together at all? Read how it stays one connected body →

What changed on 3 September 2026

Two things moved, and one of them was our own mistake.

We re-collected the cast data on 25 August 2026, which took the set from 635 titles to 666 and from 1,355 people to 1,391. And we found that this piece had joined people by name rather than by Wikidata Q-number, so two different people who share a name were counted as one. The rest of this site joins on Q-numbers precisely because Korean names repeat; this article was the exception. It is now joined the same way as everything else.

Recounted, the pairs went from 13,558 to 14,144 and the repeats from 250 to 268 — 1.9% rather than 1.8%. We also wrote the shuffle as a script this time (scripts/measure-kcw-casting-shuffle.mjs, seed 20260903) rather than running it by hand, so the null model can be re-run whenever the data moves. On the new data thirty shuffles give a median of 214 and a range of 194 to 253, and no shuffle out of thirty beat the observed 268.

The finding did not change. Repeat pairing is real, small, and close to chance — 1.3 times it rather than 1.2. There is still no repertory company here.

One limit we should have stated the first time: the shuffle keeps each title’s cast size but does not keep how many titles each person appears in. It removes the tendency for the same people to gather in the same productions, which is what we wanted to test, but it is not a null model that holds an actor’s workload fixed. A shuffle that did would be a harder test than this one.