What institutional buying before an earnings surprise tells you

On Korean provisional earnings announcements, institutional net buying in the twenty trading days before the announcement predicts the next sixty trading days of size-adjusted return negatively. The usable form is avoidance, not accumulation: among the strongest surprises, those the institutions had already crowded into underperform the rest.

Korea · provisional earnings announcements · 2011-03-18 to 2026-06-01 · the daily foreign-flow study is here

Where the effect holds, and where it does not

The effect is confined to provisional announcements, the filings that carry genuine news. Periodic reports, which mostly confirm what is already known, show nothing. Neither does foreign flow — the series this project originally set out to use.

Horizon

Nothing at five or twenty days, then the whole effect at sixty. Only three horizons were examined, so the search was narrow, but a result living at a single horizon is weaker than one that accumulates.

Why an earlier version of this page said −117.7

The retracted -117.7 bp/SD came from weighting each day equally after discarding days carrying fewer than twenty-five announcements. Both choices push the same way; the sweep below is the evidence.

Weighting each day equally sounds neutral until the days carry between one and several hundred announcements. Weighting each position equally — what a book actually earns — and clustering on the reporting season gives the figures at the top of this page.

Which investors carry it

Eight categories were searched at once, so the threshold was set by permutation before the numbers were read: shuffle the baseline within each announcement day, re-estimate the whole family, keep the largest |t|. The bar is 2.91 at the 95th percentile of 200 permutations. Pension funds clear it, and their coefficient is essentially the whole aggregate. Entered together, however, neither survives — the categories are not separable, and the aggregate remains the honest form of the signal.

The usable form is avoidance

Among the strongest surprises, those institutions had already crowded into underperform the rest. The long side does not work: the random ranking separates the quiet surprises better than the real one does, which is how a contrast that looks promising turns out to be machinery.

Alternative explanations

Each threat is one interaction on the full sample rather than a split, so the quantity of interest carries its own standard error. Index rebalancing would have to concentrate the effect in large caps, since the global and domestic benchmarks are large-cap indices; the interaction is small and points the other way. The dividend cluster and the review windows are likewise flat. The one live interaction is market-wide institutional buying, where the effect roughly triples — consistent with mechanical price pressure rather than information, though four interactions were tested and this one does not clear a family-wise threshold on its own.

Stated limitations

• The effect appears at sixty trading days and at no shorter horizon (abn5 -6.2, abn20 -1.0). Only three horizons were examined, so the search was narrow, but a result that lives at one horizon is weaker than one that builds across them.
• Half-sample stability cannot be claimed. The aggregate series reads -39.9 then -77.7 across the two halves and the pension series reads -83.1 then -48.1 -- opposite patterns, which is what thirty-four seasons buys you.
• Surprise is proxied by the announcement-day price reaction, which is entangled with the flow being measured. A true consensus-based surprise is the single largest available improvement and requires paid data.
• The pension fund cannot be claimed as a sharper baseline than the aggregate: the paired seasonal difference is -31.3 bp at t -1.47.
• The long side does not work. A random ranking separates the quiet surprises better than the real one does.

Tested and rejected

• Foreign ownership flow predicts post-event returns (+21.8 bp/SD, t +1.06)
• The effect holds across filing types (periodic filings: -19.8 bp/SD, t -1.31)
• Buying the quiet surprises earns a premium (placebo ranking reads higher)
• The pension fund is a separable carrier of the aggregate signal (both collapse when entered together)
• Index rebalancing explains the effect (size interaction +64.3, t +1.14 -- wrong sign)

Method

Sample: provisional filings. Outcome: abn60, winsorised 1/99 within the day. Regressors: surprise + c_mom20 + c_mom60 + c_size + c_vol + c_turn + baseline, rank-standardised within the day. Weighting: one vote per event. Inference: SE clustered on the reporting season. Days carrying fewer than 5 announcements are dropped. Every figure on this page is read from a single results file regenerated from the panel; none is typed by hand.

Generated 2026-09-14 03:00 UTC.