UK stock prediction (FTSE): the flashy fluke that collapsed — exactly as I warned

I run Trading Agent, and I publish every prediction my model makes — winners and losers — at /predictions. Most founders writing about their own product would lead with their best-looking number and let you assume it means something. A while back I did the opposite: I showed you our best-looking number and then spent a whole article telling you not to believe it. Today I get to do the rarest thing in this industry — come back and show you I was right to distrust it. The UK number did exactly what I said it would.

The fluke I warned you about, and what it became

Earlier this year, on UK-listed equities (the FTSE, traded on the London Stock Exchange), my model's verified directional accuracy looked spectacular — but across only a few dozen verified predictions. I wrote at the time that a gaudy figure on a sample that small was almost certainly luck wearing a nice suit, that the confidence interval was so wide it reached from "genuinely good" down to "no better than a coin," and that I expected it to regress hard toward a coin flip once the sample grew into the hundreds. I said, in bold, do not trust our UK number yet.

It regressed. Today, with the verified UK sample orders of magnitude larger, the flashy early figure is gone — it was always noise on a tiny sample — and what's left is the real, boring, trustworthy result: roughly a coin flip with a slight tilt in our favour. I'm deliberately not printing the exact figure here: a number pasted into a blog post starts rotting the day it's published. The live, always-current one — market by market, with every resolved call behind it — is on the public self-audit, and every individual call is at /predictions. This is not a disappointment. This is the system working exactly as advertised. A small edge measured over a large sample is worth infinitely more than a large edge measured over a tiny one, and now we finally have the large sample.

Why a big sample is the only number I'll actually stand behind

Here's the uncomfortable maths I walked through last time, run in reverse. Over a few dozen calls, a literal coin-flip model — 50% true accuracy, zero skill — can post a headline accuracy deep in the seventies without breaking a sweat. The random swing on a tiny sample is enormous. But that swing shrinks fast as the sample grows. Once the verified calls run into the thousands, the noise band around the true accuracy is narrow enough that what you see is very unlikely to be a fluke in either direction. It's not flattering, but it's real.

So a modest figure on thousands of rows is roughly the opposite kind of number from a spectacular one on a few dozen. The first looks unremarkable and means quite a lot; the second looks amazing and means almost nothing. What the UK's settled reading means, honestly, is a modest edge — a hair above a coin flip, not a money printer. There is one bright spot worth naming: on a 7-day horizon specifically, the UK model runs a touch stronger in the live record. Longer holding windows give a momentum-and-pattern model more room to be right, and the UK is one of the markets where that shows up. But even the bright spot is a small tilt, not a crystal ball.

Put it in context with the rest of the book. Our blended accuracy across all 16 markets hovers essentially at a coin flip — which is the honest baseline for what this kind of model does. Our strongest markets, per the live record, are Canada and the US. The UK now sits in the middle of that pack: a touch below the leaders, comfortably above the blended average, and — crucially — no longer an outlier I have to apologise for. The early spike was an outlier. Where the UK sits today is just where a real, slightly-better-than-random model lands once the data is deep enough to be honest.

This is the whole reason I built the brand around radical honesty. The industry standard is to find your flukiest, best-looking market and frame it as proof of skill — I wrote about that pattern in why most AI stock-picking tools are lying. I had that flashy figure in hand, the exact kind of number a less honest operator would have put on a billboard, and instead I told you to ignore it and watch it fall. It fell. If a tool ever shows you a gaudy accuracy figure, ask how many predictions it's measured over — and then ask to watch it for a year.

The market structure (a real footnote, not the headline)

There are genuine reasons the FTSE behaves differently from other indices, and they're worth understanding — but they're context for the model's behaviour, not an excuse for any one number.

The FTSE 100 is unusually value- and dividend-heavy, and it is one of the most globally exposed major indices in the world. Most of its large constituents — the oil majors, the miners, the global banks, the consumer-goods giants — earn the majority of their revenue abroad, not in Britain. The practical consequence is that the index often trades on the pound and on global macro as much as on the UK economy itself. A weaker pound can lift the index even on a gloomy domestic day, because overseas earnings translate back into more sterling. There's also a lingering Brexit-era re-rating overhang that has kept UK valuations structurally cheaper than US peers for years.

One mechanical detail that trips people up: UK share prices are quoted in pence (GBp), not pounds. A stock at "250" is trading at 250 pence, i.e. £2.50 — a hundredfold difference that matters enormously if you're computing returns and not paying attention. My pipeline normalises for it; a careless one wouldn't. All of that structure makes the UK a genuinely different animal for a momentum-and-pattern model like mine. But notice what it does not do: it doesn't turn a hair-above-coin-flip live record into something it isn't. Interesting structure plus an honest sample still equals a modest edge.

What I actually do with an edge this thin

The app never tells you to buy or sell anything — it shows indicators, analytics, and the scored public record — and when I mention an index-level name as an example, I'm using it to illustrate structure (globally-exposed, sterling-sensitive large-caps), not pointing at anything to trade. A directional read that sits a hair above a coin flip is real information, but it's faint information; I weight it as such, and I'd urge you to do the same. It is the kind of edge that might matter across hundreds of decisions and means nothing across one or two.

So the honest UK story has a clean arc, and you got to watch every frame of it: a great-looking early figure that I refused to trust, a public prediction that it would regress, and a settled, modest edge on a real sample that proves the caution was right. That's the entire promise of this thing — not that the number is high, but that the number is true. If you want to see exactly how these calls are generated, scored, and verified, it's all on the methodology page, and the live, unedited record — UK included — is always at /predictions. One more honest caveat: any backtest you see from me is the optimistic ceiling, because backtests get to learn from history. The live record is the real, lower number — and for the UK, the live record is the only number I'll point you to.

See the evidence for yourself — download the full resolved-prediction dataset, read the live public self-audit, 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 UK-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).