Australia stock prediction (ASX): the honest, early numbers

I'll give you the Australia number before I give you the context, because that's how this company is supposed to work. Across roughly 42 verified predictions on ASX-listed stocks, our model has scored about 43% directional accuracy. That's a little below even (50%), and below our US benchmark. Every one of those calls is public at /predictions, losses included. And ~42 verified is a small, early sample — the model retrains as more predictions verify, so the honest number should firm up over time as the log grows.

I have to flag something about that number immediately, and it cuts both ways: 42 predictions is a tiny sample. At that size, the result is preliminary. A run of bad luck or a single rough stretch of the commodity cycle could be doing a lot of the work, and so could a run of good luck if the next fifty calls go better. So the honest framing isn't "we're bad at Australia, case closed." It's "we're below our US benchmark so far, and we genuinely don't have enough data yet to be sure." I'd rather tell you that than dress up 42 rows as a verdict.

For context, the same model scores roughly 54% directional accuracy on US large caps across about 830 verified predictions, and about 46% blended across all 2,248 verified calls we've logged. So our US sample is large enough to mean something. Our Australia sample is not — yet.

Why the ASX is harder for a chart-based model

Even setting the small sample aside, there are structural reasons I'd expect Australia to have a harder time for the kind of model we run, and they're worth understanding whether or not you ever use our tool.

The headline fact about the ASX is concentration. The index is dominated by a handful of very large miners — the likes of BHP, Rio Tinto and Fortescue — and the big-four banks. (I'm naming those as structural examples of how the index is built, not as picks.) When a small number of names and two sectors drive most of the movement, "predicting Australian stocks" quietly collapses into "are you right about the resource giants and the banks this week."

And here's the problem: what moves the resource giants isn't on the price chart. It's the global commodity cycle. This is often called a two-speed market — the resource side marching to the rhythm of offshore demand for iron ore, coal and lithium, and the rest of the economy moving to a different beat entirely. The prices that actually decide whether BHP or Fortescue has a good quarter are set in offshore markets, frequently overnight while the ASX is closed. Our features are some flavour of price and volume momentum: the model reads how a stock has been trading and extrapolates the short-term tendency. A momentum model like that simply cannot see the iron-ore or lithium cycle coming. It sees yesterday's tape; the move is being decided in a different market, in a different time zone, by buyers of a physical commodity.

There's a second structural quirk that shapes ASX behaviour: franking credits, Australia's dividend-imputation system. Because franking lets domestic investors avoid double taxation on dividends, it pulls a large, income-focused investor base toward high-yielding names — the banks especially. That changes why people hold and trade those stocks. A chunk of the register is there for after-tax income, not for a short-term price view, and behaviour driven by tax structure and dividend dates doesn't show up cleanly as the momentum pattern my model is trained to detect.

Put those together and you get a market where the biggest movers are steered by a commodity cycle set offshore, and a big slice of the rest is steered by a tax-driven hunt for franked income. Neither of those is visible in a price-and-volume chart. That's a structural mismatch with our approach, independent of how lucky or unlucky those first 42 calls happened to be.

Why I'm not faking confidence to cover the gap

The easy move would be to quietly drop Australia, or keep publishing calls while hiding the hit rate behind a slick dashboard. Plenty of tools do exactly that — I wrote about the incentives in why most AI stock-picking tools are lying. The dishonest version of this business is more profitable. I'm trying to run the other one.

So let me be precise about what the Australia number does and doesn't tell you. It does not yet prove our model can't work on the ASX — the sample is too small for that claim, and I won't pretend otherwise. What it does tell me is that our current feature set is a poor structural fit for a commodity-and-franking-driven market, which is exactly what I'd predict from first principles. Closing that gap probably means inputs we don't ship today — commodity-cycle signals, offshore overnight moves, dividend-event tagging — rather than more momentum features stacked onto a market that mutes them. You can read precisely how the current model is built, and what it does and doesn't look at, on our methodology page.

The practical takeaway, stated plainly: on Australia, treat our output as still developing and preliminary. We label every call Bullish, Neutral or Bearish, never as instructions, and on this market in particular the confidence behind that label is lower and the track record is thinner. As those ~42 calls grow into a few hundred, the real picture will sharpen — and you'll watch it move on the public log, in real time, with nothing hidden. A tool that can't tell you where it's below its own benchmark and where its data is too thin isn't a research tool. It's marketing.

---

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