Japan stock prediction (Nikkei / TSE): the honest numbers and why it's hard
We score ~37% directional accuracy on Japan β below our US benchmark, on an early and still-growing sample. Here's the honest reason the Nikkei and TSE are hard for our AI model.
I'd rather show you our lowest market number so far than hide it. Across roughly 318 verified predictions on Japanese equities β the Nikkei and the broader Tokyo Stock Exchange β our model has landed about 37% directional accuracy. That is a little below even (50%) so far β on this market the model hasn't yet found a reliable edge. Worth saying up front: the Japan sample is still early and growing, and the model retrains as more predictions verify, so this honest number should firm up over time. Every one of those calls is public at /predictions, losses included, because that's the whole point of this company.
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 Japan isn't a rounding error. It's currently our lowest market number, and I find it genuinely more interesting than the markets where we do okay.
So let me be honest about why, instead of pretending it's noise.
Our edge is momentum, and Japan fights momentum
Most of our features are some flavour of price and volume momentum β the model learns from how a stock has been trading and extrapolates the short-term tendency. That has a harder time in a market where price formation is driven by things that never show up in a price chart. Japan is that market, for at least four structural reasons.
Cross-shareholdings and keiretsu ties. A large slice of Japanese share registers is still held by affiliated companies, banks and group partners β stakes held for relationship reasons, not because anyone is trading a view. When a meaningful fraction of the float doesn't move on fundamentals or sentiment, clean price discovery gets muted. The signal my model is trained to read is partly absent by design.
Lower retail participation. Historically, Japan has had thinner retail involvement than the US, and far thinner than Korea, where individual traders dominate flow. Retail activity is noisy, but it's also where a lot of the short-term momentum behaviour my features are built to detect actually lives. Less of it means less for the model to grab onto.
Years of Bank of Japan ETF buying. For a long stretch, the BOJ was a mechanical, price-insensitive buyer of Japanese equity ETFs as a matter of monetary policy. That is a giant flow that no technical indicator can anticipate β it doesn't respond to momentum, valuation, or news the way a normal participant does. A model trained to read market-driven supply and demand is, in effect, reading a tape that a central bank has been leaning on. The distortion is invisible on the chart and lethal to a momentum model.
Governance reform you can't see in the technicals. The TSE is in the middle of a serious corporate-governance push β most visibly the campaign pressuring companies trading below book value to fix it, the so-called "PBR above 1" effort. When a re-rating is driven by policy and management reform rather than by trading behaviour, it arrives as a step-change that the price history simply doesn't foreshadow. My features see the past; this kind of move is about a future the past doesn't contain.
Why I'm not "fixing" this by faking confidence
The easy move would be to quietly drop Japan, or to keep publishing calls while burying the hit rate. Plenty of tools do exactly that β I wrote about the incentives in why most AI stock-picking tools are lying. The dishonest version of this business is very profitable. I'm trying to run the other one.
A below-even number tells me something real: our feature set is mismatched to the regime, not that the market is unbeatable. Beating Japan probably requires inputs we don't yet have β flow and ownership data, policy-event tagging, governance-reform signals β rather than more momentum features layered on top of a market that mutes momentum. That's an honest research problem, and it's a different model than the one we ship today. You can read exactly how the current one is built, and what it does and doesn't look at, on our methodology page.
Here's the practical takeaway, stated plainly: on Japan, weight your own research more heavily here. We label every call Bullish, Neutral, or Bearish, never as instructions, and on this market in particular the confidence behind that label is low. I'd rather you knew that than discovered it the expensive way. A tool that can't tell you where it's still developing isn't a research tool β it's marketing. We publish the 37% so you can hold us to it.
This article is educational content about machine learning and market structure. It is not financial advice, not a recommendation to buy or sell any Japanese-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).
