StockDriftsStockDriftsWe studied 3+ years of Korea's DART insider disclosures — 170,720 share-count changes across 56,930 D002 filings and 944 D005 pre-disclosure plans, from 16,350 individual insiders at 2,705 Korean companies. Using cross-sectional alpha at 21-, 63-, and 126-day fixed forward horizons, we measured which Korean insider signals actually predict future stock performance.
Korea's insider-disclosure system is unusually informative compared to the US or Japan. Two filings stand out: D002, which records every share-count change by directors, officers, and 10%+ holders within five business days — open-market trades, RSU grants, gifts, mergers, the lot — and D005, a Korea-specific pre-trade disclosure where insiders publicly announce intent to buy or sell before executing.
No major equity market — not the US, Japan, or Europe — requires insiders to disclose their trading intentions before they trade. Korea does. Every D005 plan filing is a public commitment, and every withdrawal of one is itself a signal.
Combined with mandatory same-day reporting on actual trades, Korea's 2024 Value-Up program, and the ongoing chaebol governance reform pressure, this creates a uniquely rich disclosure dataset for measuring insider conviction — and one informative bearish signal that doesn't exist in other markets.
The biggest single cohort effect in our data, in either direction: when an insider withdraws a planned sell, the stock drops 22% in six months on average. It earns its own alarm tier (and a Tier C placement in the table below).
An insider files an intent to sell, then withdraws the plan before executing. The story: the insider learned material bad news after filing, and pulling the plan avoids regulatory scrutiny. The data agrees.
For comparison: the average D005 sell filing (not withdrawn) drifts only −1.7% over six months. The withdrawal itself adds a 20-percentage-point bearish overlay. We treat this as its own dashboard alarm tier — separate from the standard S/A/B/C structure.
Not every share-count change is an insider decision. The discretionary-decision filter matters more than magnitude.
Korean DART tags every insider event with an acquisition method (취득방법) — about 30 distinct codes covering open-market trades, RSU grants, mergers, gifts, IPO disclosures, option exercises, and more. Not all of them represent an insider acting on private information.
A real insider signal requires three things: the trade was a discretionary decision, made by the insider personally, with their own capital. These three filters drop a surprising amount of the dataset.
One cohort in our data has tier-S magnitude but fails the discretionary-decision test. We exclude it from the insider buying feed:
Several “BUY”-tagged events stay persistently tier-C — they're mechanical share-count increases tied to corporate actions and consistently underperform: 출자전환 (debt-for-equity conversion, −20% α), 행사가액조정 (strike-price adjustment), 주식매수청구권 행사 (appraisal-rights exercise), 전환사채의권리행사 (CB conversion), and 신규등록 (new registration). The dashboard hides these from the “insider buying” feed and shows them only as risk-flag annotations.
Notably, several signals that looked tier-C in our 26-month sample — 합병 (merger), 신규상장 (post-IPO disclosure), CB인수 (CB acquisition) — washed out to tier B once we extended to 3 years. This is why methodology requires multi-regime data: a single bear year can make any corporate action look bearish.
The raw dataset is 170,720 share-count changes from 56,930 D002 filings — DART's ownership-change disclosures covering open-market trades, RSU grants, gifts, mergers, and other share-count changes by registered insiders and 10%+ holders — plus 944 D005 pre-disclosure plans, filed between January 2022 and April 2026 by 16,350 individual insiders at 2,705 listed Korean companies. For alpha measurement we collapse each filing to one event per insider and require matched daily price history; that yields 51,041 events in the forward-return universe used for the tier table below.
Korea's 3-year window spans multiple regimes: the 2022 rate-hike drawdown (KOSPI −25%), the 2023 chip-cycle rebound, the 2024 Value-Up rally, and the 2025 bull run. Raw forward returns would be dominated by which year a trade fell in. We solve this with cross-sectional alpha: for every trade, we subtract the same-day median forward return across all 2,705 stocks. What's left is “did this specific stock beat the average Korean stock?” — regime-free by construction.
Returns are clipped to ±50pp to dampen lottery-winner effects. We report three horizons: 21 trading days (~1 month), 63 days (~3 months), and 126 days (~6 months). Tier thresholds are set on the 126-day horizon, the longest with adequate data coverage.
Korea's Financial Supervisory Service runs DART (Data Analysis, Retrieval and Transfer system) — the public registry of all corporate disclosures. Two report types matter for insider monitoring:
Filed within 5 business days of any change in shareholding by directors, officers, or 10%+ holders. Discloses the insider, position, transaction date, acquisition method, share count change, and resulting ownership %.
A Korea-specific filing requiring insiders to publicly announce planned trades before executing them. Includes planned shares, price basis, and trading-window dates. Withdrawals and amendments are public events too — and themselves predictive, as the alarm above shows.
Cross-sectional α at fixed horizons — every filing measured against the same-day universe median.
126-day alpha ≥ +15% AND win rate ≥ 60%. Once corporate-action artifacts are excluded, only the 5-filter composite qualifies.
| Signal | 6 Month α | Win Rate | Filings |
|---|---|---|---|
Composite — Open-market BUY + cluster ≥ 6 + non-major + 10M-1B + mom > 0 Five conviction filters stacked on 장내매수 — the highest-conviction recipe in the dataset | +30.07% | 86.4% | 339 |
126-day alpha ≥ +5% AND win rate ≥ 53%. Solid positive alpha — workhorse open-market buys, outside-director buys, D005 BUY plans, multi-insider clusters.
| Signal | 6 Month α | Win Rate | Filings |
|---|---|---|---|
장내매수 + cluster ≥ 6 + non-major shareholder ≥6 distinct insiders buying open-market in 30 days, none a 10%+ holder | +14.50% | 61.6% | 1,666 |
상속:BUY (Inheritance acquisition) Insider acquired shares via inheritance — small but persistent edge across 3y | +14.54% | 71.1% | 45 |
장내매수 + 사외이사 (Outside Director BUY) Outside directors have least access to compensation programs — almost always discretionary | +14.43% | 71.0% | 124 |
D005:BUY (Pre-disclosure BUY plan) Insider publicly filed intent to buy before executing | +10.66% | 62.8% | 191 |
교환:SELL (Share exchange / swap) Insider 'sold' via corporate share-swap — remaining stock outperforms over 6 months | +10.50% | 64.2% | 53 |
장내매수:SELL (Open-market 'BUY method' tagged as SELL) Data quirk — net position decreased on a buy-method leg; predictive of stock rally | +9.95% | 60.5% | 372 |
신규선임:BUY (Newly appointed executive) Disclosure on new executive appointment with shares — n nearly tripled with 3y data | +9.92% | 60.9% | 3,152 |
장내매수:BUY (Open-market BUY — workhorse) The default conviction signal — open-market discretionary buy, n=12,259 over 3 years | +8.35% | 58.3% | 12,259 |
신규보고:BUY (New 5%+ report) Initial 5%+ shareholder filing — the Korean equivalent of a US 13D | +7.71% | 57.0% | 1,005 |
126-day alpha within ±5%. No predictive value — most open-market sells, option exercises, after-hours block trades, and several corporate-action methods that washed out as the dataset grew.
| Signal | 6 Month α | Win Rate | Filings |
|---|---|---|---|
장내매도:SELL (Open-market SELL) The default sell — neutral as a stand-alone signal | -0.07% | 44.1% | 3,834 |
주식매수선택권:BUY (Stock-option exercise) Mechanical option exercise — not a discretionary buy | -2.07% | 38.6% | 1,604 |
D005:SELL (Pre-disclosure SELL plan) Mild bearish drift but not enough to flag standalone | -1.74% | 44.3% | 228 |
장외매도:SELL (Off-market negotiated SELL) Block trades to specific buyers — borderline tier-C | -4.37% | 35.8% | 689 |
126-day alpha ≤ −5%. The D005 sell-withdrawal alarm, debt-for-equity conversions, strike-price adjustments, and forced/mechanical corporate-action 'buys'.
| Signal | 6 Month α | Win Rate | Filings |
|---|---|---|---|
D005:SELL_WITHDRAW ⚠️ (Cancelled SELL plan) Strongest single bearish signal — insider planned to sell, then withdrew | -22.16% | 31.6% | 19 |
출자전환:BUY (Debt-for-equity conversion) Creditor takes equity in exchange for forgiving debt — extreme distress signal | -19.75% | 13.0% | 46 |
행사가액조정:SELL (Strike-price adjustment SELL) Mechanical adjustment with bearish realized direction | -11.45% | 30.0% | 50 |
신규등록:BUY (New registration disclosure) Initial registration disclosure — usually post-IPO or restructuring | -8.84% | 40.0% | 85 |
행사가액조정:BUY (Strike-price adjustment BUY) Mechanical — not a buy decision | -7.03% | 34.0% | 424 |
주식매수청구권 행사:BUY (Appraisal-rights exercise) Dissenting shareholders selling back to the company — bearish on remaining stub | -6.89% | 36.3% | 559 |
전환사채의권리행사:BUY (CB rights exercise) Forced/structured acquisition via CB conversion — bearish on average | -6.86% | 34.3% | 73 |
The single highest-conviction recipe in our research.
No single acquisition-method × transaction-type cell makes tier S once we exclude the corporate-compensation artifacts. The genuine tier-S signal lives in a five-filter composite:
Each filter is independently meaningful. Open-market enforces the discretionary-decision test. Cluster ≥ 6 means six or more different insiders independently decided to buy the same stock within 30 days — Bayesian aggregation across multiple private signals. Excluding major shareholders strips out defensive control-block buying. The KRW band excludes both token-size disclosures and large strategic blocks. Positive momentum confirms the broader market hasn't already repudiated the thesis.
This analysis is for informational and educational purposes only. Past performance does not guarantee future results. Signal tiers are based on historical cross-sectional alpha and should not be used as the sole basis for investment decisions. Several cohorts (notably D005 sub-types and small-sample acquisition methods) have fewer than 100 events and should be treated as directional indications. We will revise tier assignments as more data accumulates.
See these signals live on the Korea Insider dashboard.
Open Korea Insider Dashboard