Asia Pulse
Research/Methodology

Six hypotheses we rejected in Korean biotech

July 26, 2026
SIX REJECTED · ONE SURVIVED
Key numbers
6 / 8hypotheses rejected
185events tested across three catalyst types
3–15×standard deviation over the mean

Our US-to-Korea transmission coefficient works in semiconductors. It does not work here. We tested eight hypotheses and rejected six — and the reason turned out to be the mechanism itself.

We measure how movement in US markets transmits into Korean equities, and we express it as a coefficient. In semiconductors it holds. When Micron moves, SK Hynix follows: beta 0.379, R² 0.165, n = 1,632, walk-forward validated.

So we assumed biotech would work too. Same country, same exchange, same kind of headline-driven catalyst. A drug gets approved, a trial reads out, the stock reprices.

It did not work. Of eight hypotheses, six were rejected.

What we tested

HypothesisSampleResult
FDA approval events72noise
Phase 3 trial readouts38noise
Phase 1 trial readouts75noise
Small-cap amplificationR² 0.006
Nvidia → AI drug discoveryweaker than control
Power-sector coefficient3 namesR² 0.026

Why “noise” is a measurement, not a shrug.

For each event we measured beta-adjusted excess returns against the index across a window around the event date. In every segment the standard deviation ran three to fifteen times the mean.

At that dispersion the average describes nothing. Half the sample goes one way, half goes the other, and the midpoint is an artefact. Publishing “average move of X%” from data like this would be worse than publishing nothing.

0% FDA approval Phase 3 Phase 1 n = 72 n = 38 n = 75 band = ±1σ · block = mean
Schematic. The mean sits inside a band several times its own width, in all three event types. A number that small inside a band that wide is not a signal.

The hypotheses, one by one

Small-cap amplification

If market capitalisation is small, a single trial outcome should represent a larger share of enterprise value, so the reaction should be bigger. R² 0.006, tested on two listed small-cap names. Effectively no explanatory power.

Nvidia to AI drug discovery

We expected Korean AI-drug-discovery names to track Nvidia. They tracked it less closely than the KRX300 IT control did. The hypothesis was not merely unproven — it pointed the wrong way.

Power-sector coefficient

Average R² of 0.026 across three names, below the KOSPI benchmark of 0.134. A sector coefficient that explains less than the index it sits inside is not a sector coefficient.

Semiconductors KOSPI benchmark GE Vernova → power Healthcare Power (3-name avg) Small-cap biotech 0.165 0.134 0.104 0.040 0.026 0.006 Solid = survived validation · Faded = rejected
R² by hypothesis. The KOSPI benchmark is the line everything else has to clear — two sit above it, four do not.

Why it failed

Our first assumption was that we lacked data. That was wrong.

Published event-study research on pharmaceutical stocks already describes what we found. For large diversified pharmaceutical companies the impact of any single event tends to be smaller, and for late-stage biotechnology companies clinical trial results move the stock less than one might expect.

All six of our names sit in exactly that category — large-cap KOSPI biotech, the kind that runs dozens of programmes at once.

Celltrion runs 44 concurrent clinical trials. Hanmi runs 62. One readout does not re-rate a company running sixty of them. When a pipeline is that diffuse, individual events are absorbed rather than priced.

So the signal was not missing. The diffusion was the mechanism.

How we use a rejection

A negative result is still a product. This is close to what our chain analysis actually says when the biotech axis comes up:

We measured three event types and found none of them move these names. That is not a failure of the data — it may be the mechanism. A company running 44 concurrent trials does not re-rate on any one of them. That same diffusion is what a defensive holder is buying.

This is a reading, not a finding.

The last line matters. We do not sell an interpretation as a verified result. Whatever we publish carries the status of the claim alongside the claim.

Rules this produced

A number without a control group means nothing. Semiconductor R² 0.165 is only meaningful next to healthcare at 0.040 — a 4.2× gap is what licenses the phrase “semiconductor-specific.” In isolation, no coefficient reads as strong or weak.

If sigma exceeds twice the mean, do not ship the mean. An average is information only when the dispersion is small enough for it to describe something.

Block post-selection bias. Back-testing the drugs that made headlines is circular reasoning. Pull the full set or pull nothing.

A weak coefficient can still carry a product, if there is a ledger. Our defense axis has R² 0.066 — weak. But the confirmed record says $66.6M in US federal procurement, all of it maintenance, zero new-build. The coefficient was not the story. The ledger was.

What is still open

Two of the eight hypotheses have not been rejected, and one exception surfaced in the power axis: GE Vernova alone showed R² 0.104, specific enough to be interesting. Constellation Energy, in the same sector, came in at 0.026. When the underlying business differs, transmission does not carry.

That sample is thin. We are still watching it, and we will publish the outcome either way.


A rejection is not a failure. Knowing what does not work is worth roughly what knowing what does work is worth — particularly if you say it out loud.

Asia Pulse publishes Korea equity chain analysis with the coefficients, the ledgers and the falsifier conditions attached. Six endpoints, priced per request, no API key.

See the endpoints →
Discover more
Research
Method notes, every other day
What we measure, what we reject, and the conditions under which we would be proven wrong.
Products
Korea equity chain, priced per call
Transmission coefficients, customs and procurement ledgers, falsifier checks. Settled on-chain.