NZX AI forecast: can machine learning predict New Zealand stocks?
Every global AI stock tool ignores the NZX — too small, too illiquid, not worth the engineering. As a New Zealand company, we don't. Here's an honest look at whether machine learning can forecast NZX-listed stocks — including the uncomfortable fact that, on roughly 1,823 verified predictions, our home market currently runs below a coin flip.
I built Trading Agent in Auckland. So the question I get asked most by other Kiwis is the simplest one: does any of this work on the NZX?
It's a fair question, and an underserved one. Search for "NZX AI forecast" or "New Zealand stock prediction" and you'll find almost nothing — because every global AI-stock tool has made the same business decision: New Zealand is too small to bother engineering for. The NZX's entire market cap is a rounding error next to a single mega-cap on the Nasdaq.
We made the opposite decision. And because we did, I can give you something almost no one else will: a large, fully public, timestamped track record on our home market — and an honest reckoning with what it says. Spoiler: the number is not flattering.
The NZX is a genuinely different animal
Before any numbers, you need to understand why the NZX is harder to model than the S&P 500 — because the reasons are structural, not just "less data."
1. Liquidity
Many NZX names trade thinly. A handful of large caps — the likes of Fisher & Paykel Healthcare, Auckland International Airport, Mainfreight, Spark, Meridian — carry most of the liquidity, and below them the order books get thin fast. Thin liquidity means wider spreads, larger gaps, and price moves driven by single large orders rather than the smooth, continuous flow a technical model assumes. A model trained on the deep, liquid US tape sees a much noisier signal here.
2. A small, low-turnover, correlated index
The NZX 50 is dominated by a few sectors — utilities, healthcare, property, and a heavy weighting toward defensive, dividend-paying names. That makes the index behave differently from a growth-heavy US market, and it means individual-stock moves are often driven by interest-rate expectations and dividend dynamics rather than the momentum and volatility patterns technical features capture best. It's the same structural problem we see in Malaysia: a small, low-turnover market gives a momentum model very little clean signal to work with.
3. Offshore-driven flow
A large share of NZX activity is institutional and offshore. When a global fund rebalances its Australasian allocation, an NZX large-cap can move on flow that has nothing to do with the company or its chart. Those moves are, almost by definition, unpredictable from price history alone.
4. Data and corporate actions
Free data feeds handle NZX corporate actions — imputation credits, frequent capital raises, dividend reinvestment — less cleanly than US data. As with Taiwan, some apparent model error is really data-hygiene noise we're still working through.
The honest number: below a coin flip
Here's where I have to be straight with you in a way most tools never are.
As of writing, our public verified log holds roughly 1,823 verified NZX predictions, running at about 48.0% directional accuracy. Read that again: 48.0% is below 50. On our own home market, over a sample large enough that the number actually means something, our model is currently a touch worse than a coin flip on direction.
I'm not going to dress that up. A few-dozen-prediction sample could be noise; 1,823 cannot be waved away. This is a real, sustained, sub-even result, and the structural reasons above are exactly why I'd expect a small, thin, offshore-driven market to be brutal for a momentum model. New Zealand is hard for the same reasons Malaysia is hard — and Malaysia is the other market where we sit below even.
So the honest answer to "does ML work on the NZX?" is: right now, on our live record, no — not as a directional edge. I'd rather tell you that to your face than bury it. Most tools would either hide this market or quietly quote you a backtest. We're doing neither.
Backtests lie a little; the live number is the truth
This is the part worth tattooing on the inside of your eyelids. In a backtest, almost any model looks decent on the NZX — you can curve-fit a clean-looking history out of thin liquidity if you try. Backtests are optimistic by construction. The honest figure is the live, forward, timestamped one, and that's the lower one: ~48.0%. Whenever a backtest and a live record disagree, believe the live record.
How the NZX compares to the rest
We run 16 markets: the US, Canada, the UK, Taiwan, China, Singapore, Malaysia, Vietnam, Australia, New Zealand, Korea, Japan, India, Hong Kong, Indonesia, and Thailand. (Yes — India is fully covered, not "restricted." So is everything else on that list.)
Blended across all sixteen, our verified directional accuracy sits around 49.7% — itself basically a coin flip, which is the whole reason we publish every market's number instead of one flattering headline. The spread within that average is the real story:
- Best: Canada and the US, both around 53% — small but genuine, on large samples.
- Worst: small, thin markets like New Zealand (~48.0%) and Malaysia, which drag below even.
The contrast is the point. A number is only as trustworthy as the sample behind it and the market behind it, and we show you both on every market — including the ones where we're losing.
Why bother with the NZX at all?
If our home-market number is below even, why keep covering it?
- Honesty means showing the losses, not just the wins. A "radical honesty" tool that quietly dropped its worst market would be a fraud. The NZX number is exactly the kind of thing most tools hide, which is exactly why we don't.
- Somebody should try. Kiwi retail investors — and there are a lot of them now, post-Sharesies — deserve a research tool that at least tries their market honestly and tells them plainly when it isn't working, instead of pretending it doesn't exist.
- It's our home market. A tool built in New Zealand that ignored the NZX would be a bit embarrassing. A tool built here that lied about the NZX would be worse.
What I'd actually tell a Kiwi investor
Don't lean on the NZX number as a signal right now — full stop. On the live record it's below a coin flip, so treat it as honest evidence of a hard market, not as a forecast to act on.
What the model gives you on the NZX is a research read, not a recommendation — the technical and macro signals it weighs, and an honest, sample-backed measure of how often it has actually been right — never a "buy," a "sell," a price target, or a guaranteed percentage. On a market where our verified hit rate is ~48.0%, the most useful thing that record can tell you is "weight your own research far more heavily than this." That's a legitimate, honest use of a model that refuses to pretend it's an oracle.
For the markets where our model currently has a real, sample-backed edge, look to US and Canadian large-caps — see the methodology and the full per-market breakdown at /predictions. And if the NZX number ever climbs above even as the model improves, you'll watch it happen live, with every win and loss on the record.
That's the deal: no hype, no hidden losses, and an honest "this one's below a coin flip" when that's the truth.
See the evidence for yourself — download the full resolved-prediction dataset, read the live public self-audit (hit-rate confidence intervals, live-vs-backfill split), inspect every model card, or run the research tools on your own data. No hype, just the receipts.
This article is educational content about machine learning and market structure. It is not financial advice, not a recommendation to buy or sell any NZX-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).
