aikstockdata/korea-equity-daily

Korean Equity Daily Prices + DART Filing Impact

Daily settled closes for 1,463 Korean listed companies (KOSPI and KOSDAQ) over the last 250 trading days, plus a table of what stocks did after each type of regulatory filing.

Korean equity data is oddly hard to get. The official sources are free but gated: the Financial Services Commission open-data portal wants an API key and returns raw payloads with Korean field names, and DART (the disclosure system) is a separate registration. Commercial APIs either paywall Korea or cover only large caps. This is the same data, already joined, as plain CSV.

from datasets import load_dataset

px = load_dataset("aikstockdata/korea-equity-daily", "daily_prices", split="train")
print(px[0])
# {'code': '000020', 'name_ko': '동화약품', 'market': 'KOSPI',
#  'date': '20250725', 'close_krw': 6780, 'volume': 34614}

Contents

daily_prices — 361,143 rows

One row per stock per trading day, 2025-07-25 → 2026-08-04 (250 trading days).

column	type	note
code	string	6-digit Korean ticker, zero-padded. Keep it a string — 005930 is Samsung Electronics
name_ko	string	Company name in Korean
market	string	KOSPI (634 names) or KOSDAQ (829)
date	string	YYYYMMDD, Korea Standard Time trading date
close_krw	int	Settled close in won. Not adjusted for splits or dividends
volume	int	Shares traded. 0 means no trades, which is a fact, not a gap

Holidays and non-trading days have no row. Do not assume date continuity — align on the dates present, not on a calendar range. Newly listed names have fewer than 250 rows.

Identifiers are strings, on purpose. code, date, rcept_no and every date field are strings. A Korean ticker is six digits including leading zeros — 000020 is Dongwha Pharm, and 20 is nothing at all. Read as integers, they stop joining to anything, including this project's own per-stock files at /data/public/s/000020.json.

This is why the loadable files here are JSON Lines, not CSV. CSV carries values without types, so the reader guesses — and the guess destroys the zero padding. It was wrong on this dataset until it was caught and fixed. The .csv files are still in the repository because they are a third of the size and convenient over curl, but they are not what the configs above load. If you read the CSVs yourself, force the identifier columns to text:

pd.read_csv("daily_prices.csv", dtype={"code": str, "date": str})

stocks — 1,463 rows

Master list as of the 2026-08-04 close — the same trading day daily_prices ends on: code, name_ko, market, close_krw, change_pct, market_cap_krw.

filing_impact_summary — 22 rows

The one people come for. Every DART filing in the collection window is joined to that company's own daily closes and to its own index, giving the median market-adjusted return after each filing type. 644 filings so far.

Filing type	+1 trading day	+5 trading days
Supply contract	-0.18% * / 46% / n=142	-0.27% * / 48% / n=81
Preliminary earnings (consolidated)	-0.14% * / 49% / n=140	+0.49% * / 51% / n=51
Preliminary earnings (separate)	-1.50% * / 39% / n=74	+1.32% * / 59% / n=22
Dividend decision	-0.16% * / 49% / n=47	+2.21% * / 69% / n=32
Largest shareholder change	+1.53% * / 67% / n=27	n=15 (withheld)
Treasury stock trust contract	+2.04% * / 70% / n=20	n=13 (withheld)
Periodic financial report	-0.46% * / 45% / n=20	n=9 (withheld)

median market-adjusted return / share that beat its index / distinct price paths. * marks a 95% interval that spans zero — that median is not distinguishable from zero. Fifteen more filing types are in the file with samples too small to report.

Read that asterisk before anything else. Across all four horizons there are 18 cells carrying a number and 17 of them are starred. Exactly one is not: treasury-stock trust contracts on the baseline day (+2.26%, interval +0.57 to +4.09, beat rate 80% with an interval of 58–92%, n=20). Everything else in this table is a number you cannot distinguish from zero. Publishing the intervals was the point; almost none of these survive them.

Adding six days of data made the table weaker, not stronger. On 2026-08-05 two cells cleared zero; on 2026-08-06, with 644 filings instead of 598, only one does. Dividend decisions at +5 days fell from +3.52% (interval +0.23 to +8.93) to +2.21% with an interval that now spans zero. Periodic financial reports came back into the table at +1 day by crossing the n=20 threshold — at -0.46%, starred. This is what a small sample looks like from the inside, and it is the reason the intervals are printed rather than the medians alone. The numbers here are regenerated from the live site on every upload; expect them to move again.

An earlier revision fixed a double-count. Until 2026-08-05 each DART receipt number counted as one observation. When one company files three documents on the same day the baseline and the entire price path are identical across them, so a single company-day was counted three times. Aggregation is now keyed on issuer × baseline date × filing type. Both counts are published: n is distinct price paths, n_filings is receipts.

filing_price_impact — 644 rows

The individual filings behind that table, one row each, so you can recompute the aggregates instead of taking them on trust — and disagree with them.

column	note
rcept_no	DART receipt number. dart_url opens the original document
code, name_ko, market	the filing company
filing_type_ko, filing_type_en, kind	type, matching the summary table exactly
cluster	issuer + baseline date. Rows sharing one are the same price path — dedupe on this before aggregating, or you will double-count
price_break	true where an ex-rights date or share consolidation resets the quoted price
base_date, base_close_krw	the baseline: first trading-day close on or after receipt
h0_*	the baseline day itself — previous close into the baseline close
h1_*, h5_*, h20_*	horizon date, raw return, index return, and the difference

import statistics
from datasets import load_dataset

ev = load_dataset("aikstockdata/korea-equity-daily", "filing_price_impact", split="train")
seen = {}                       # 같은 issuer-day 는 한 번만 — 집계와 같은 규칙
for r in ev:
    if r["filing_type_en"] == "Dividend decision" and r["h5_excess_pct"] is not None:
        seen.setdefault(r["cluster"], r["h5_excess_pct"])
x = list(seen.values())
print(round(statistics.median(x), 2), len(x))
# 2.21 32  — the same number the summary reports

Every published median reproduces exactly from these rows; that is checked before each upload.

Method. Baseline is the first trading-day close on or after the filing receipt date. Returns are measured at +1, +5 and +20 trading days — five rows forward in that stock's own series, not a calendar offset, because Korean market closures are irregular. From each return the stock's own index (KOSPI or KOSDAQ) over the identical window is subtracted. The statistic is the median, not the mean. Amended filings are dropped because they duplicate the original.

Every published median ships with a 95% interval. *_median_ci95_lo/hi come from order statistics, *_up_ratio_ci95_lo/hi from a Wilson score interval — both closed-form and deterministic, so they are identical on every rebuild. *_ci_includes_zero is set when the interval spans zero, which means that median is not distinguishable from zero. A bold number without its interval is the dishonest option; 70% at n=20 has a Wilson interval of 48–86%.

The h0 columns are the baseline day itself, measured from the previous trading day's close. DART accepts filings during the session and after it, and the receipt time is not in the public data, so the two cannot be separated. A filing made mid-session is already partly reflected in that day's close, and that part disappears into the baseline. h0 exists to make the missing piece visible, not to remove it.

Each horizon covers a different set of filings — *_base_date_from/to say which. A longer horizon excludes recent filings that have not had time to elapse, so its sample clusters earlier. Reading two horizons side by side as "what happened N days later" is wrong when those windows differ.

When one company files several documents on the same day, they share a baseline and therefore an identical price path. Those count once per filing type; n_filings records how many receipts sat behind that count.

Fewer than 20 observations gets no number at all — a median over eight cases turns coincidence into a statistic. That is why the _enough column exists and why most of this table is withheld.

Four types are withheld no matter how large the sample gets: paid-in capital increase, bonus issue, paid-in and bonus issue, and 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 a market reaction. Their individual rows stay in filing_price_impact, flagged price_break — not hidden, just never averaged. The +20 day columns are entirely empty and left visible: collection started 2026-07-20, so twenty trading days have not elapsed for anything yet.

This is a record, not a claim. A filing and a price move inside the same window does not mean one caused the other. Earnings, sector rotation and the market itself are all in there. It is not a signal and it is not investment advice. The task_categories tags on this card say what the data can be used for in a search index; they are not a claim that anything here forecasts anything.

What is not here

Read this before integrating, so you can stop early if it matters:

• No real-time or intraday prices. These are previous-trading-day settled closes from the government feed (T+1). If you need live quotes, this is the wrong dataset.
• No PER, PBR, target prices, analyst ratings, investor-type flows, or sector tags. Those are brokerage-derived; only public-sector data is redistributed here.
• No adjustment for splits, mergers or dividends. Closes are as-reported.
• Korean equities only, and only the ~1,463 names with a quote on a normal trading day — not the entire listed market. Suspended, delisted and just-listed names fall out.
• The filing sample is young. Every median above will move.
• Company names are Korean. Column names are English.

Snapshot vs. live

main is overwritten on every upload — it is not citable. This repository is regenerated from the live site each time it is refreshed, so a number you quote from main may not exist here next week. The table above already moved once between two consecutive uploads.

To cite, pin the revision. Every upload is a commit, and commits are permanent:

load_dataset("aikstockdata/korea-equity-daily", "filing_impact_summary",
             revision="<commit sha from the repo history>", split="train")

Quote that revision together with the generated_kst stamp carried inside the JSON.

Two other ways to get a number that stays put:

• A frozen monthly repository. aikstockdata/korea-equity-daily-YYYY-MM is uploaded once and never updated. Cite it by name; no revision hash needed.
• A dated file on the site. Every publish leaves an immutable copy at https://aikstockdata.com/data/public/snapshots/disclosure_impact_YYYY-MM-DD.json, listed in snapshots/index.json. Aggregates only, no per-filing rows, but it is the same table.

The pipeline republishes every trading evening; that live version, with no signup, no API key, no rate limit and CORS open, is at:

• Catalog of every file with byte size and freshness: https://aikstockdata.com/data/public/index.json
• One stock's year of closes as a single ~7 KB file: https://aikstockdata.com/data/public/s/005930_history.json
• Filing impact, in English, with the method written out: https://aikstockdata.com/en/filing-impact
• The same as JSON, per-filing values keyed by DART receipt number: https://aikstockdata.com/data/public/disclosure_impact.json (aggregates only, smaller: disclosure_impact_summary.json)
• OpenAPI 3.1 spec: https://aikstockdata.com/openapi.json
• English overview: https://aikstockdata.com/en

One caveat worth stating: the CDN's bot filter returns 403 to the default Python-urllib user agent. requests, curl, httpx and browser fetch all work; with urllib, send any User-Agent header.

Source, license, citation

Derived from Financial Supervisory Service DART (disclosures) and the Financial Services Commission open-data portal (settled quotes), both Korean public-sector data. Redistribution and commercial use are permitted with attribution. Real-time quote redistribution is not permitted and is not done here.

Data: aikstockdata.com — source: FSS DART, FSC Korea Open Data Portal

Disclaimer

Information only. Not investment advice, and not a recommendation to buy or sell any security. Every number is a record of the past. Investment decisions and their consequences are the reader's own.

---

한국어 요약

한국 상장기업 1,463종목의 최근 250거래일 확정 종가·거래량(361,143행)과, DART 공시 유형별로 접수일 이후 시장조정 수익률 중앙값을 정리한 표입니다.

• daily_prices — 종목×거래일 종가·거래량. 휴장일은 행이 없으므로 날짜 연속성을 가정하지 마세요. 수정주가 아님(액면분할·배당 미조정).
• stocks — 2026-08-04 종가 기준 종목 마스터(daily_prices 의 마지막 거래일과 같음).
• filing_impact_summary — 공시 22유형. 기준점은 접수일 이후 첫 거래일 종가, 구간은 거래일 기준, 소속 시장 지수 등락률을 뺀 값의 중앙값. 표본 20건 미만은 수치를 내지 않습니다. +20거래일 칸은 아직 전부 비어 있습니다(수집 시작 2026-07-20).

인과가 아니라 기록입니다. 같은 구간에 공시와 주가 변동이 함께 있었다는 사실일 뿐이며, 투자 권유가 아닙니다.

이 저장소는 2026-08-05 스냅샷입니다. 매 거래일 갱신되는 원본은 가입·API 키·요청 제한 없이 https://aikstockdata.com/data/public/index.json 에서 받습니다.

출처: 금융감독원 DART · 금융위원회 공공데이터포털 — 공공데이터 가공물로 출처 표기 후 영리 목적 포함 자유 이용.