US stock prediction with machine learning: where our model actually works

I run Trading Agent, and I publish every prediction we make — including the ones that lose — at /predictions. So when I tell you the United States is the one market where our model genuinely earns its keep, I'm not selling you a dream. I'm telling you what roughly 3,076 verified predictions actually show, and then I'm going to spend the rest of this article explaining why "actually works" is a much smaller claim than most of the industry wants you to believe.

The number, and what it isn't

On US-listed equities, our model's verified directional accuracy is about 52.6% across roughly 3,076 verified predictions. That's our deepest sample and, alongside Canada, our strongest market — both sit at about 53%. For context, our blended accuracy across all 16 markets we cover is around 49.7% — essentially a coin flip. So the US genuinely stands apart, even if the gap is measured in a couple of points, not a chasm.

One more honest wrinkle worth surfacing: accuracy isn't uniform across horizons. On the 7-day horizon specifically, US accuracy is closer to 54% — a touch better than the all-horizon blend. I mention that not to cherry-pick the prettier number, but because it's true and you should know the full shape of the data, not just the headline.

Now let me kill the fantasy before it forms. 52.6% is not a money printer. It is a modest edge. If you flip a fair coin you get 50%, so we're talking about a couple of percentage points of signal above noise. That sounds underwhelming, and honestly it should — but a small, real edge applied consistently across many independent decisions, with disciplined position sizing, is worth something, whereas a big, fake edge applied once is worth nothing. The whole game is telling those two apart. I wrote about why so much of the industry blurs that line in why most AI stock-picking tools are lying, and it's the backbone of how I think about this.

Why the US is where it works

This isn't luck, and it isn't because US stocks are magic. It's market structure.

The US is the deepest and most liquid equity market on earth. That liquidity matters enormously for a model like ours, because the features we lean on — RSI, MACD, moving averages, volume, volatility — are all derived from price and volume behaviour. In a thin, sparsely-traded market, those signals are mostly noise: a couple of large orders can swing the tape and your "momentum signal" is really just one fund rebalancing. In the US, continuous, liquid order flow means a momentum or volatility pattern is far more likely to reflect something structural rather than an accident of low volume.

There's also the training-data angle, which people underrate. The US market has deep, clean, long price history. When you train a machine-learning model, the quality and length of your historical data is doing a lot of the quiet heavy lifting. We had more to learn from here, and it shows. Frankly, our feature set was effectively tuned against US-style data — so it's not a coincidence this is the market where it performs best. We optimised on home turf.

The flip side of the same coin: the US is institution-dominated and fairly efficient. Smart, well-resourced money is competing on every tick, which is precisely why the edge is small. An efficient market doesn't hand out large, durable edges to a solo-built model — if it did, the edge would be arbitraged away. So the 52.6% isn't despite the market's efficiency; it's the natural ceiling that efficiency imposes. A small real edge is roughly what an honest model should find in a market this competitive. If someone shows you 80% on US large-caps, your first instinct should be suspicion, not envy.

A word on backtests versus the live record

When we develop and tune the model, backtests on historical data tend to look better than reality — that's the nature of fitting against the past. The number that matters is the one earned forward, in public, on predictions made before the outcome was known. That live public record is the real number, and it's lower than any backtest. The 52.6% above is the live figure, not the flattering lab one. I'd rather you anchor on the honest, lower number than a polished one I can't stand behind.

How to actually use a 52.6% signal

If you take one thing from this, take this: a modest edge is an input, not an instruction.

We publish Bullish / Neutral / Bearish directional reads on large-cap names — and I want to be precise about what those are. They are model output describing a probability lean, not a "buy" or "sell". I'm not going to tell you to buy any specific stock, here or anywhere, because (a) I'm legally not in that business, and (b) it would be intellectually dishonest given what 52.6% actually means. The right mental model is: this is one signal you might weigh alongside your own research, your risk tolerance, position sizing, fundamentals, and whatever else you bring to a decision. Compounded over many independent decisions, with disciplined sizing, a slight directional tilt can be useful. Bet the farm on any single call and the edge is far too thin to protect you.

I'd also gently point out that mentioning a ticker like AAPL or MSFT in this article is me using them as examples of large-cap market structure — the kind of liquid, well-covered names where our signals are least noisy. It is not me pointing at them. The structural point is the point.

The honest bottom line

The US is where our model has a real, sample-backed, structurally-explainable edge — and it's still only about 52.6%. I'd rather tell you that plainly than dress it up. If you want to see the full methodology behind how we generate and verify these calls, it's at /methodology, and the running scoreboard of hits and misses is always at /predictions. Read both before you decide what, if anything, our signal is worth to you. That's the deal with radical honesty: I show you the work, and you make up your own mind.

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 US-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).