Research
한국어로 읽기 →Tests of common trading beliefs against actual market data. Sample sizes, periods and test methods are stated, and negative results (rejections) are published as they are. Conclusions go only as far as the data does.
Part 5 — Second-Round Tests
Trading on Earnings Day — Announcement Premium, PEAD, and Pre-Announcement Prediction: Three Hypotheses Tested on 34,504 Events
We pre-registered and tested three hypotheses on 34,504 US large-cap earnings announcements (2021–2026). The announcement premium was +0.2%, exactly equal to transaction costs; post-positive-surprise drift existed only in 2023–24; and public information available before the announcement could not predict the direction of the reaction. The only result consistent across all three sub-periods was the 3-month underperformance of negative-surprise stocks — material for an avoidance filter, not a buy signal.
When the US Falls, So Does Korea — A Lead Correlation of 0.34, and Why It Can't Be Traded
Across 4,318 trading days from 2010 to 2026, the correlation between the S&P500's prior-day return and KOSPI's same-day return is +0.336 (t 23.4). Aligned on the same date it is +0.170, half as large. On the day after the US falls -1%, KOSPI averages -0.81% and declines 71% of the time. But this relationship is already reflected in the opening gap, so it is not tradable information.
Insider Cluster Re-Test — We Extended It to 45 Quarters, and Found a Look-Ahead Error in the Process
[Correction included] We extended Part 12's U.S. insider cluster buying to 45 quarters of the SEC dataset (2015Q1–2026Q1). The day after publishing the extension result (60-day +4.04%p), we found that the backtest entered at the closing price on the filing date itself — a look-ahead error. On the actually tradable next-session open, 20-day is +0.57%p (t 2.0), and the 60-day effect was created by the single year 2020. Removed from the live-candidate list.
Do Ten Weak Signals Combine into a Strong One? — The Premise Behind IR = IC × √Breadth
If you combine signals with individual ICs of 0.01–0.03 and increase the number of holdings and the rebalance frequency, do risk-adjusted returns improve? On the S&P500 from 2011 to 2026, all ten signals had ICs below 0.06, and low volatility, the lottery effect and the 52-week high carried the opposite sign from the literature. In the holdout (S&P400+600), the strongest signal (revenue growth, t 1.5) flipped to t -2.6.
Risk Rules Don't Escape Post-Hoc Fitting Either — A 189-Month Test of Correlation and Volatility Caps
The rule "exclude a stock if its correlation with an already-picked name is 0.5 or higher," created right after the July crash, turned that July from -18.5% into -8.0%. Tested on the 158 months before the tuning, IR goes 1.38→1.47 (within noise), MDD -29%→-28%, and cumulative return is halved. The large drawdowns were beta crashes, which a correlation cap never touches.
Part 4 — What Survived
Why Only Momentum and Quality Remain — Series Wrap-Up
After testing dozens of hypotheses, what remained were two factors academia had already validated over decades. The condition for a factor to survive is that it must be painful to hold — and an individual investor's real edge lies not in factors but in size, patience, and events.
Small Survivors — Two Narrow Signals That Survived 16 Years of Testing
Out of dozens of hypotheses, only two Korean-market signals escaped rejection. Buying down days during extreme KOSPI slumps (60% win rate, but only 2.4 opportunities a year) and the rights-offering new-share listing calendar (+4.5%/56%, but ±40% dispersion). Whether either is "large enough to use" is a separate question.
The Only Copy-Trading Strategy That Passed — U.S. Insider Cluster Buying
[Correction 2026-09-04] The original +2.78%p (t 5.6) reflects a look-ahead error in which the backtest entered at the closing price on the filing date itself. Using the next trading day's open — the price you could actually buy at — the same sample gives +1.22%p (t 2.6), and the full period (2015–2026) gives +0.57%p (t 2.0). Three-quarters of the filing reaction was already over before the open.
Part 3 — Self-Autopsy
Five Ways Backtests Deceive You — A Collection of Actual Cases We Experienced
A spike stock strategy that looked like +9.4%p due to look-ahead bias, sample distortion caused by an API returning only 30 results, and an accident where all stocks were classified as scarce stocks because we missed the unit of a thousand shares — these are not someone else's theories, but 5 types of backtest traps we actually fell into.
The Momentum Strategy's Invoice — July 2026, a -32.5% Month in a Live Account
A record of the July 2026 crash (-32.5%, versus KOSPI -20.6% in the same month) taken while running a momentum strategy in a live account. A high-beta stock's 477% upside capture ratio also works on the way down, at 210%. This pain is exactly why the momentum premium does not disappear.
How We Fooled Ourselves With Our Own Backtest — Autopsy of a Post-Hoc Theme Sleeve
A theme sleeve showed a +272%p improvement in total return over a 12-year backtest, so we deployed it live. Three months later we found that the improvement was essentially one year — 2023 — and that the sector list itself had been chosen after the fact. We removed it. A self-autopsy.
Part 2 — Charts and Spikes
Buying Stocks After a Crash — 50,000 Deep-Drawdown Purchases Tested
Buying names that fell 15% or more over five days returned -3.8%p versus the market after 20 days, beating the market only 39.4% of the time (n=49,694). Neither buying the bounce after a capitulation nor betting on a rebound conditioned on investor flows on down days showed any edge.
Do Candlestick Patterns and Double Bottoms Work — Testing 5 Textbook Patterns
Four candlestick reversal patterns — shooting star, dark cloud cover, engulfing, and long bearish candle — showed no difference in subsequent returns from the baseline. The 42 KOSPI double-bottom neckline breakouts had a 55% 60-day win rate, actually lower than buying on any random day (59%).
Buying Before the Spike — Five Front-Running Signals, 820,000 Cases Tested
Instead of chasing spikes, we tested five strategies that enter early when the precursors appear (foreign/institutional accumulation, volume surges, joint price-and-volume increases), across 820,000 samples. All were zero or negative versus the market. The worst was "quiet rises with no identifiable buyer" at -5.47%p.
What Happens If You Chase Spike Stocks — All 22,474 Cases Tested
Buy a stock at the close after it spiked +13% or more that day, and 20 days later you are at -7.5%p versus the market (median -8.9%), with a beat-the-market rate of 34.8%. Add any condition — news ignition, chart pattern, market cap — and every path was negative.
Part 1 — Following Others
Buying What Buffett Buys — A 17-Quarter Replication Test of the Berkshire Portfolio
Replicating the top-10 holdings from Berkshire Hathaway's 13F filings on each filing date returned +55.0% cumulative from 2022-08 to 2026-08. SPY returned +89.7% over the same period. Quarterly win rate 7/16, average quarterly excess -1.32%p.
Copying U.S. Congress Members' Stock Buys — 4,095 Disclosures Tested
Buying alongside U.S. House members' stock purchase disclosures returned -1.47%p versus SPY after 120 trading days (t=-3.1). It significantly underperformed the market. The Pelosi tracker's high returns are better explained by call-option leverage than by copying stocks.
Following Stocks That Institutions and Private Equity Funds Are Accumulating — Tested with Ranking Correlation IC
The correlation (IC) between monthly net-buying rankings of institutions and private equity funds and next-month return rankings averaged +0.010 over 43 months — effectively zero. Chasing stocks that spiked after five days of concentrated institutional buying returned -4.1%p versus the market over 20 days.
Does Following Foreign Investors' Net Purchases Yield Returns? — A 4-Year, 986-Trading Day Test in the Korean Market
Buying the top 20 stocks with the highest 5-day net purchases by foreign investors yielded +1.81%p over the market after 20 days. However, stocks heavily sold by foreign investors also rose by +1.74%p. The long-short difference was +0.03%p — the buy direction itself contained no information.
This article documents tests on historical data for informational purposes only. It is not investment advice or a recommendation to buy or sell any security. Past test results do not guarantee future returns.