TW stock prediction with machine learning: the honest numbers and why it's hard

If you trade Taiwan-listed stocks — TSMC (2330), MediaTek (2454), Hon Hai (2317) — you've probably noticed that almost every "AI stock prediction" product on the internet quietly skips your market. They cover US large-caps, maybe a few European names, and stop.

We don't skip Taiwan. We run the same machine-learning pipeline on TWSE-listed large-caps that we run on the S&P 500. But I'm not going to pretend it's where our model is strongest, because it isn't — yet. This is the honest version.

One thing to keep in mind up front: the Taiwan sample is still early and growing. The model retrains as more predictions verify, so the honest number you see here should firm up over time as the data accumulates.

The number: ~41% directional accuracy on TW

As of writing, across ~310 verified Taiwan predictions in our public log, our directional accuracy on TW large-caps is about 41%.

That is a little below even (50%).

Compare that to our US large-cap number — about 54% on ~830 verified predictions — and you can see the gap. The exact same model architecture, the same features (RSI, MACD, moving averages, volume, volatility), the same walk-forward validation, produces a genuinely useful edge on US large-caps and a noticeably weaker result than the US on Taiwan.

Why? This is the interesting part, and it tells you something real about how markets work.

Why Taiwan is harder to predict

1. Retail-dominated order flow

Taiwan's market has one of the highest retail-participation rates in the world — historically well over half of daily turnover comes from individual investors, versus a much more institution-dominated US market. Retail-heavy flow tends to be noisier and more sentiment-driven on short horizons. The technical patterns a model learns from price and volume are weaker signals when the marginal trader is reacting to a LINE group rather than a discounted-cash-flow model.

2. Day-trading mechanics and tax incentives

Taiwan has actively encouraged day trading through reduced transaction-tax rates on same-day round trips. That concentrates a lot of activity into intraday mean-reversion that washes out by the close — which is exactly the kind of move a daily-horizon model has a harder time anticipating. The signal the model is trained to find (continuation) and the behaviour the market actually exhibits (intraday reversion) are working against each other.

3. Concentration and correlation

A huge share of the TWSE's market cap and movement is TSMC and the semiconductor supply chain. When one name and one sector dominate, a lot of "individual stock" prediction collapses into "are you right about TSMC and the global chip cycle this week" — which is driven by US customer demand, foreign fund flows, and geopolitics, none of which live in a price chart.

4. Data quality and corporate-action noise

yfinance and other free data sources handle TWSE corporate actions (the frequent ex-dividend and capital-reduction events common in Taiwan) less cleanly than US data. Some of the model's TW weakness is almost certainly noise injected by imperfect adjusted-close handling, not pure model failure. We're investigating how much of the 41% is fixable data hygiene versus genuine market efficiency.

What this teaches you

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

If a simple technical model lands a little below even on Taiwan large-caps, that's partial evidence the market is reasonably efficient on short horizons — the easy patterns have been arbitraged away by all those active retail and prop traders. That's not a bug in the market; it's the market doing its job.

It also means you should be deeply suspicious of any tool that claims high accuracy on Taiwan stocks. If our honest, walk-forward-validated number is 41%, and someone else is advertising "89% accuracy on 台股," 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.

So why cover Taiwan at all?

Three reasons:

1. The model output is still useful as context, not as a signal. A Neutral read with an honest 41% historical accuracy tells you "this is a market where our patterns don't hold — trust your own thesis here, not ours." That's a legitimate, honest use.
2. The other tools available to TW retail investors are technical-indicator dashboards with no honesty layer at all. A research tool that shows 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 — Taiwan, Japan, Korea, Vietnam, Malaysia, alongside US, UK, EU, Australia, Canada and our home NZX. Covering them honestly, including where we're weak, is the whole point.

If you want the markets where our model is currently strong, that's US large-caps — see the methodology page and the per-market breakdown at /predictions. And if our Taiwan number improves as we fix the data handling, you'll see it move on that same public log, in real time, with nothing hidden.

---

This article is educational content about machine learning and market structure. It is not financial advice, not a recommendation to buy or sell any Taiwan-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).