Canada stock prediction (TSX): promising, but too small a sample to trust
Our model hits ~56% on the TSX — above a coin flip — but on only ~36 verified calls. That's far too small to trust, and here's why sample size matters.
I run Trading Agent, and I publish every prediction we make — winners and losers alike — at /predictions. So I could open this article by telling you that our model scores about 56% directional accuracy on the Canadian market (the TSX), which is comfortably above a coin flip and, on its face, one of our better numbers. That would be true. It would also be close to dishonest if I stopped there, because that 56% rests on only about 36 verified predictions. Thirty-six. That is not enough data to claim anything, and the rest of this article is about why I'm telling you to distrust my own good-looking number.
The number, and the asterisk that swallows it
Let me put it next to our other markets so the problem is obvious. On US large-caps, our verified directional accuracy is roughly 54% across about 830 verified predictions. Our blended accuracy across all 13 markets is around 46% on roughly 2,248 verified calls. The US number and the Canada number look similar — 54% versus 56% — but they are not remotely the same kind of claim. One is built on 830 independent data points; the other on 36.
Here's the intuition. Flip a fair coin 36 times and you will fairly often get 20 or 21 heads — that's 56–58% — purely by chance. Nobody would call that a biased coin. The swing you'd expect from sheer luck on 36 trials is wide enough to produce a "56%" with no underlying skill at all. So when our TSX accuracy reads 56%, the honest interpretation isn't "we have an edge in Canada." It's "we can't yet tell whether we have an edge in Canada, and the same data is perfectly consistent with the true rate being 45%." As the sample grows toward a few hundred calls, I fully expect that headline number to move — possibly down, maybe a lot. I'd be unsurprised to see it settle below 50%.
This is the single most abused trick in this industry, and I wrote about it at length in why most AI stock-picking tools are lying: show a great hit rate, hide the denominator. A 56% on 36 rows and a 56% on 3,600 rows look identical in a marketing screenshot and mean completely different things. The denominator is the whole story.
Why the early read might not be a fluke (a hypothesis, not a victory)
Now I'll argue the other side honestly, because there is a structural reason the Canada number could turn out to be real once we have the data to confirm it.
The TSX is resource-heavy — energy and materials dominate the index in a way that makes it structurally cousin to Australia's market. On its own, that doesn't obviously help us. But Canada has a second feature that does plausibly matter: a large share of its biggest names are dual-listed on US exchanges. Companies in banking, energy, and mining trade simultaneously in Toronto and New York, which means a meaningful chunk of Canadian large-cap price formation inherits US-style liquidity and order flow. The price you see on the TSX is being arbitraged tick-by-tick against a deep, continuous US tape.
That matters because the US is, by a clear margin, our strongest market — and the reason is precisely that liquidity and clean price discovery let our momentum and volatility features read signal instead of noise. If a slice of the Canadian market is effectively riding on US-grade liquidity, it's reasonable to hypothesise that our US-tuned feature set travels north better than it travels to thinner, more idiosyncratic markets. That is a genuinely plausible mechanism. It is also just a hypothesis. A believable story for why a number might be real is not the same as the number being real — you still need the sample to prove it, and we don't have it yet.
How I'm treating Canada right now
Cautiously optimistic, strictly disciplined. I am not going to claim an edge on 36 data points, and you shouldn't grant me one. We publish Bullish, Neutral, and Bearish directional reads on Canadian large-cap names — and those are descriptions of a model's probability lean, never instructions to buy or sell. On a market this thinly sampled, the right posture is to treat the output as a weak, provisional signal at most: one input you might weigh against your own research and risk tolerance, not something to lean on.
When I mention that many TSX names are dual-listed, I'm pointing at market structure as the interesting variable — the liquidity those names inherit — not at any particular company. The structural point is the point.
You can see exactly how we generate and verify every call on our methodology page, and the live scoreboard — Canada included, sample size and all — is always at /predictions. Watch that 36 grow. If the 56% holds up over a few hundred calls, then we'll have something worth talking about. Until then, it's a promising line on a chart with far too few points behind it, and I'd rather tell you that than sell you the headline.
This article is educational content about machine learning and market structure. It is not financial advice, not a recommendation to buy or sell any Canada-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).
