Korea stock prediction (KOSPI): our honest numbers on a retail-driven market

If you trade KOSPI-listed stocks — Samsung Electronics (005930), SK Hynix (000660), the big chaebol names — you've probably noticed that most global "AI stock prediction" tools either ignore Korea entirely or quietly bundle it into some "Asia" bucket they never really tuned for.

We run Korea. The same machine-learning pipeline we run on the S&P 500 — RSI, MACD, moving averages, volume, volatility, all walk-forward validated — runs on KOSPI large-caps too. But I'm going to give you the honest version, because the honest version is the whole brand: every prediction we've ever made, wins and losses, is public at /predictions.

The number: ~46% directional accuracy on Korea

As of writing, across ~209 verified Korea predictions, our directional accuracy on KOSPI large-caps is about 46%.

That is, to put it plainly, close to even — a slight bias below 50%, in fact.

Compare that with our US large-cap number — about 54% on ~830 verified predictions — and against our blended cross-market number of about 46% on ~2,248 verified predictions. The exact same model architecture that produces a small but real edge on US large-caps produces essentially no edge on Korea.

A note on the sample: 209 is moderate. It's big enough to take seriously — this isn't a 20-prediction fluke I'm hand-waving over — but it's not so large that I'd treat ~46% as a precision figure. Call it "close to even (a slight bias below 50%), and we have enough data to be fairly confident it's close to even." And the model keeps retraining as more Korea predictions verify, so the number should firm up over time. That nuance matters, and most tools won't give it to you.

So why is Korea so hard? This is the genuinely interesting part, and it tells you something real about how this market works.

Why KOSPI is close to random for a technical model

1. The "ants" drive the flow

Korea has one of the most retail-dominated equity markets in the world. Individual investors — affectionately called 개미 (gaemi, "ants") — account for a very large share of daily turnover. Ant flow is fast, emotional, and herd-prone: it can pile into a name and back out again on news, a YouTube tip, or a KakaoTalk rumour, with little to do with fundamentals.

For a model that learns from price and volume, that's a problem. The technical "patterns" the model is trained to find assume the marginal trader is at least loosely anchored to something persistent. When the marginal trader is a swarm of ants reacting to sentiment on a daily horizon, continuation signals break down and short-term moves look a lot like noise — because, in an information sense, they often are.

2. Extreme index concentration

KOSPI is dominated to an unusual degree by Samsung Electronics and a handful of other chaebol names. When one stock and one supply chain carry that much of the index's weight and movement, a lot of "individual stock prediction" quietly collapses into one question: are you right about Samsung and the global memory-chip cycle this quarter?

And that question isn't answered by a price chart. It's driven by global DRAM and NAND demand, foreign fund flows, US export policy, and the capex cycles of Samsung's customers. None of that lives in the RSI.

3. Short-selling bans keep changing the rules

Korea's regulators have periodically imposed outright short-selling bans — sometimes market-wide, sometimes lifted and reimposed — as a policy response to volatility. Whatever you think of the policy, the modelling consequence is brutal: a short-selling ban abruptly changes market microstructure. The mechanics of price discovery before a ban, during a ban, and after it's lifted are simply not the same regime.

A model trained across those periods is trying to learn one stable relationship from data generated under several different rule sets. That structural instability is exactly the kind of thing that drags a directional edge back toward 50%.

4. High single-stock volatility

Korean single-stock volatility runs hot, and the daily price-limit structure plus sentiment-driven swings produce sharp, gappy moves. High volatility without a persistent, learnable driver doesn't help a directional model — it just widens the distribution of outcomes around a coin flip.

What this teaches you

Here's the lesson I value more than any single prediction: a market being hard to predict is itself information.

If a reasonable technical model can't beat a coin flip on KOSPI large-caps over 209 calls, that's partial evidence the easy short-horizon patterns have been arbitraged away — washed out by all that active retail flow and the regime shifts from interventions. That's not a flaw in the market. That's the market being efficient in the ways that matter for a price-only model.

It also means you should be deeply suspicious of anyone advertising high accuracy on Korean stocks. If our honest, walk-forward-validated number is ~46%, and someone is selling "90% accuracy on KOSPI," ask them the question from our piece on AI stock-picking tools: show me the complete, timestamped, loss-included log. They won't have one, because a real one looks like ours — and ours, on Korea, sits close to even — a slight bias, honestly disclosed.

So why cover Korea at all?

Three reasons.

1. An honest Neutral is still useful. A read labelled with its true ~46% historical accuracy tells you something real: this is a market where our patterns don't hold — weight your own thesis here, not ours. We report Bullish / Neutral / Bearish, never Buy / Sell, and on Korea the honest meta-signal is "don't lean on us."
2. The alternative tools offer no honesty layer at all. Most products serving Korean retail are technical-indicator dashboards that never show you when they're wrong. A research tool that openly publishes its own weakness on your market is more useful than one that hides it.
3. We're a New Zealand company built for the markets the US tools ignore — Korea, Taiwan, Japan, Vietnam, Malaysia, alongside US, UK, EU, Australia, Canada and our home NZX. Covering them honestly, including where we're near random, is the entire point.

If you want the markets where our model currently has a real, sample-backed edge, that's US large-caps — see the methodology page and the full per-market breakdown at /predictions. And if our Korea number ever moves, you'll see it move there, in real time, with every win and loss on the record and nothing hidden.

That's the deal: no hype, no buried losses, and an honest "this one's basically a coin flip" when that's the truth.

---

This article is educational content about machine learning and market structure. It is not financial advice, not a recommendation to buy or sell any Korea-listed or other security, and not directed at any individual's circumstances. Trading Agent is a quantitative research tool operated by WU Capital Limited (New Zealand).