Two methodology questions on the filing-impact numbers
The filing_impact_summary config is the part of this dataset I am least sure about, so I would rather have it argued with than downloaded quietly.
What it does: every DART filing in the collection window is joined to that company's own daily closes. Baseline is the first trading-day close on or after the receipt date. Returns are taken at +1, +5 and +20 trading days — rows forward in that stock's own series, not a calendar offset, because Korean market closures are irregular. From each return I subtract the stock's own index (KOSPI or KOSDAQ) over the identical window, then take the median per filing type. Fewer than 20 observations gets no number at all.
from datasets import load_dataset
s = load_dataset("aikstockdata/korea-equity-daily", "filing_impact_summary", split="train")
for r in s:
if r["h5_enough"]:
print(r["filing_type_en"], r["h5_median_excess_pct"], r["h5_n"])
Preliminary earnings (consolidated) 2.08 54
Supply contract -0.53 66
Dividend decision 3.91 27
- Is the stock's own index the right benchmark?
Right now a KOSDAQ microcap is compared against all of KOSDAQ. That index is dominated by a handful of large names, so the "market-adjusted" return of a small illiquid stock may be mostly measuring a size factor I never intended to include. The obvious alternative is matching on size (and maybe sector) before subtracting, but with 598 events the matched buckets get thin fast, and I do not have a sector classification in this data yet — it would have to come from somewhere else, which breaks the "public-sector data only" rule the rest of the project follows.
Is a crude own-index adjustment better than nothing here, or actively misleading? If you have seen a benchmark choice that survives small samples, I would like to read it.
- Same issuer, same day, several filings.
A company that files three documents on one date produces three events with the same baseline and therefore three identical price paths. They currently count as three observations. That inflates n and lets one company's move show up repeatedly inside a single filing type.
Two ways out that I can see: collapse same-issuer-same-day filings into one event (but then which filing type does the event belong to?), or keep them and cluster at the issuer-date level when aggregating. I do not know which is standard practice for event studies at this granularity.
Things I already know are wrong or missing, so you do not have to report them:
The +20 day columns are entirely empty. Collection started 2026-07-20, so twenty trading days have not elapsed for anything yet. I left the columns visible rather than hiding them.
The sample is young and every median will move.
Prices are not adjusted for splits or dividends.
Per-filing rows are not in this dataset yet. I found while building it that my own published JSON kept only the receipt number per event and dropped the filing type, so the aggregates could not be recomputed from the individual rows — a real defect in the thing I consider this project's centrepiece. It is fixed upstream and the per-filing file lands here shortly, along with a check that the summary reproduces exactly from the individual rows.
This is a record of what happened, not a claim about what causes what, and not investment advice. If the method is wrong I would rather fix the file than defend it.
Answering my own question 2. It was a real defect, it is fixed, and fixing it plus publishing confidence intervals removed almost everything this dataset had to say.
The counting was wrong. When one company files several documents on the same day, every one of them shares a baseline close and therefore an identical price path. I was counting each DART receipt number as an observation, so a single company-day entered the median several times. In the current 644 filings there are only 536 distinct price paths. One company on 2026-07-31 contributed seven receipt numbers carrying the same value; three financial holding companies contributed five or six each on 2026-07-23.
Aggregation is now keyed on issuer × baseline date × filing type. Filings of different types on the same day are not merged — I can't attribute the move to one of them, but counting it once inside each type is still a fact about that type. Same type, same issuer, same day collapses to one. Both counts ship, so you can see the gap yourself:
from datasets import load_dataset
s = load_dataset("aikstockdata/korea-equity-daily", "filing_impact_summary", split="train")
for r in s:
if r["h1_n"] != r["h1_n_filings"]:
print(r["filing_type_en"], r["h1_n"], "of", r["h1_n_filings"])
Supply contract 142 of 153
Preliminary earnings (consolidated) 140 of 158
Preliminary earnings (separate) 74 of 78
Dividend decision 47 of 51
Largest shareholder change 27 of 29
Periodic financial report 20 of 28
filing_price_impact gained a cluster column (issuer + baseline date). Dedupe on it before aggregating, or you will double-count exactly the way I did.
Now the part that matters more. Several of you said the medians meant nothing without a statement of uncertainty. Every published cell now carries a 95% interval — median by order statistics, beat rate by Wilson score, both closed-form so the numbers don't drift between publishes. The result:
Across all four horizons there are 18 cells carrying a number, and 17 of them have an interval that spans zero.
Exactly one clears it: treasury-stock trust contracts on the baseline day, +2.26% with an interval of +0.57 to +4.09, beat rate 80% with an interval of 58–92%, n=20. That is the whole list.
The one result I had been leading with is gone. Dividend decisions at +5 trading days:
2026-08-05 +3.52% interval +0.23 to +8.93 n=26
2026-08-06 +2.21% interval -2.39 to +8.75 n=32
Six more days of filings, six more observations, and the interval got wider and crossed zero. That is not a bug; that is what n=32 looks like when you stop hiding it behind a point estimate. Periodic financial reports came back into the table over the same six days by crossing the 20-observation floor — and arrived starred, at -0.46%.
I'm leaving the table published rather than pulling it, because "we measured this and it did not separate from zero" is a result, and because a dataset that only shows you its survivors is worse than one that shows you the whole distribution of attempts. But the card now says the count out loud, above the table, and I'd rather you quote the count than the medians.
Four types are withheld permanently, not for sample size: paid-in capital increase, bonus issue, paid-in and bonus issue, reverse stock split. An ex-rights date or a share consolidation resets the quoted price mechanically, and these closes are not adjusted for corporate actions, so a window containing that date measures the break rather than any market reaction. All four are currently under 20 observations, so nothing was being published — but they would have started publishing wrong numbers on their own once the sample grew. Their rows stay in the dataset flagged price_break; they are simply never averaged. The real fix is adjusted prices, and the exclusion list goes away when that lands.
Question 1 is still open and I now think it's the only one left that's interesting. Subtracting the stock's own index assumes beta = 1. Several of you pointed out that this is a real limitation but roughly the only adjustment this sample size supports, and that the honest move is to stop calling it "market-adjusted" and name the assumption instead. I agree and the card does that now. If you think matched benchmarking is workable at n=536, I'd like to hear how — and I'd take a concrete second question too: should different-type filings from the same issuer on the same day be excluded as confounded, rather than counted once inside each type? Right now they aren't, and I can't tell whether that's caution or laziness.
Numbers above are v5, regenerated from the live site on 2026-08-06. They will move again.