Pre-event institutional positioning and the earnings surprise →
Korea publishes foreign ownership per stock every day. The premise was that foreign investors digest information more slowly and split larger orders, so their flow is persistent, and that a daily disclosure of it is hard to replicate. This page is generated from the measurement outputs, so every figure below is the one the code last produced.
Some results are older than the raw data or missing: combination.json, level_se.json. Re-run those measurements before relying on this page.
Two hypotheses fit the persistence curve equally well: that foreign flow is special, or that all large institutional flow is persistent because big orders get split over days. The domestic-institution series separates them.
| lag | Institutions | t | Foreign | t | ratio f/k | ||
|---|---|---|---|---|---|---|---|
| 1 | +0.2982 | +235.9 | +0.1295 | +95.2 | 0.43 | ||
| 2 | +0.2241 | +199.6 | +0.1195 | +100.7 | 0.53 | ||
| 3 | +0.1801 | +164.5 | +0.0991 | +95.5 | 0.55 | ||
| 5 | +0.1376 | +127.4 | +0.0780 | +80.9 | 0.57 | ||
| 10 | +0.0951 | +96.9 | +0.0527 | +61.0 | 0.55 | ||
| 20 | +0.0603 | +62.2 | +0.0284 | +35.3 | 0.47 |
| horizon | mean IC | t | days | gross/yr | turnover | net/yr |
|---|---|---|---|---|---|---|
| 1d | +0.00392 | +3.93 | 3,363 | +6.25% | 320.0x | -89.73% |
| 5d | -0.00036 | -0.37 | 3,360 | -0.54% | 64.0x | -19.73% |
| 20d | -0.00405 | -4.51 | 3,350 | -1.34% | 16.0x | -6.14% |
Net is after 30bp one-way, applied to the measured turnover of both legs.
| horizon | Thin third | Middle | Liquid third | |||
|---|---|---|---|---|---|---|
| 1d | +0.00919t +7.26 | +0.00695t +7.36 | +0.00066t +0.69 | |||
| 5d | +0.00598t +3.96 | +0.00337t +3.30 | -0.00416t -4.17 | |||
| 20d | -0.00090t -0.48 | -0.00179t -1.54 | -0.00492t -4.44 |
Newey-West t beneath each coefficient. A cell greyed out is indistinguishable from zero.
| horizon | raw IC | t | residual IC | t | kept |
|---|---|---|---|---|---|
| 1d | +0.00359 | +4.02 | +0.00229 | +3.43 | 64% |
| 5d | -0.00053 | -0.60 | -0.00175 | -2.60 | 333% |
| 20d | -0.00464 | -5.16 | -0.00597 | -8.05 | 129% |
Residualised each day against size, liquidity, reversal_5d, momentum_60d, momentum_120d, volatility_20d.
| signal | IC | t as published | t Newey-West | lag-1 AC | independent obs. | reading |
|---|---|---|---|---|---|---|
| level, 20d | +0.13307 | +47.59 | +13.60 | +0.945 | 131of 1,607 | holds up |
| level_z, 20d | +0.01549 | +9.61 | +2.78 | +0.910 | 135of 1,607 | marginal |
| d60, 20d | +0.01002 | +7.57 | +2.29 | +0.887 | 147of 1,607 | marginal |
| d120, 20d | +0.01651 | +10.97 | +3.14 | +0.917 | 132of 1,607 | holds up |
| level, 60d | +0.17406 | +66.34 | +10.87 | +0.975 | 42of 1,567 | holds up |
| level_z, 60d | +0.01405 | +7.93 | +1.28 | +0.957 | 41of 1,567 | WAS significant, is NOT |
| d60, 60d | +0.01307 | +9.51 | +1.68 | +0.936 | 49of 1,567 | WAS significant, is NOT |
| d120, 60d | +0.01612 | +9.36 | +1.48 | +0.967 | 39of 1,567 | WAS significant, is NOT |
A high lag-1 autocorrelation means consecutive daily coefficients are nearly the same number, so the published t counts the same evidence many times over.
| day | vs baseline |flow| | ||
|---|---|---|---|
| -5 | 1.06× | ||
| -4 | 1.02× | ||
| -3 | 1.03× | ||
| -2 | 1.05× | ||
| -1 | 1.05× | ||
| +0 | 1.02× | ← effective date | |
| +1 | 1.07× | ||
| +2 | 1.04× | ||
| +3 | 1.03× | ||
| +4 | 1.01× | ||
| +5 | 1.02× |
136 rule-derived effective dates. A calendar that found real events would show a clear hump at zero.
The signal is a stock measure — the daily change in foreign shares held. The exchange separately publishes foreign net buying, a flow measure. Across 4,095 overlapping days and 3,615 tickers the mean daily rank correlation is +0.1046, the tenth percentile is -0.0037 and the worst day is -0.2352.
One of the two is not measuring what we assume. Candidates: settlement-date against trade-date convention, securities lending and custody transfers moving holdings without a trade, depositary-receipt conversions, in-kind index creation, and a known misclassification of orders routed through domestic brokers. It does not change the verdict — there is no edge on either reading — but it changes what the numbers mean.