AI Signals — Weekend Read: The stock that's always on sale

Weekend Read #11 — July 12, 2026

AI Investor Barometer tracks how five LLMs form DCF assumptions for 24 listed companies — daily, independently, from identical inputs.

On Thursday, July 9, 2026, Nvidia closed at $202.78. That same evening, the mean price target across 58 analysts stood at $301.62. Between the price of the world's most closely watched stock and the professional consensus sat 49 percent of implied upside — and this was no forgotten industrial nobody covers, but the company followed by more analysts than any other in our panel.

And Nvidia is no outlier. That same day, 20 of the 24 companies we track had a mean analyst target above the market price. Microsoft was promised 46 percent of upside, Meta 31, KONE 26. The median promise across the whole panel was nine percent. The only company whose target sat clearly below its price was Fortum.

Take price targets literally and the stock market is in permanent clearance: almost every stock, almost every day, available below its "true" value. Some of this is to be expected. Stocks are supposed to deliver returns, and over a twelve-month horizon a promised gain of a few percent is little more than the market's long-run base rate written down as a number. But 49 percent is not a base rate. And a base rate does not explain what research has documented across decades: the promises exceed the outcomes systematically, not randomly.

This essay is a user's manual. What a price target actually measures, why it points up almost by construction — and why it is still useful, provided you read it correctly.

A forecast that follows its subject

On paper, a price target is a simple thing: an analyst's estimate of where a stock will trade twelve months from now. Behind it usually sits a model — discounted cash flows or peer multiples — and behind the model, a set of assumptions about growth, profitability and risk. The same structure our AI models work through every morning.

In practice, a price target behaves unlike a forecast. Brav and Lehavy (Journal of Finance, 2003) worked through hundreds of thousands of US price targets and found two persistent features. First, targets sat on average 28 percent above the market price — several times the typical annual return of the equity market. Second — and this is the more revealing finding — the ratio of target to price stayed stable over time. When prices rose, targets rose after them; when prices fell, targets fell too, keeping their distance. The premium does not close. It travels with the price.

A forecast that updates to track its subject at a fixed distance is not a forecast in the ordinary sense of the word. It is closer to a convention: an established way of expressing a view on a company as a number that looks like a price.

Why does the convention point upward? The literature offers no single answer, and this essay does not need to pick one — the observation itself is enough: the premium has been documented across decades and market cycles, and it is structural rather than random error in individual estimates. The more useful question for a reader is how often the promise is kept. That, too, has been measured.

Thirty-eight percent

How often is the promise kept, then? Bradshaw, Brown and Huang (*Review of Accounting Studies*, 2013) measured it across ten years of data, 2000–2009. The returns implied by price targets exceeded realized returns by 15 percentage points on average. At the twelve-month deadline, only 38 percent of targets had been met. Count every moment along the way — the price touched the target on at least one day — and the share rises to 64 percent.

Those two numbers are best read together, because they reveal what a price target actually is. It is not an expected value around which outcomes fall symmetrically. It is closer to a favorable scenario: a defensible outcome if things go well. The price visits it more often than it stays — as befits a scenario.

Read this way, Nvidia's $301 does not mean "the stock will rise 49 percent." It means: "if demand, margins and multiples develop favorably, this level can be justified." That claim is far more modest than the number makes it look — and that is precisely why the number's format misleads. A scenario dressed up as a price stops looking like a scenario.

Optimism carries a passport

So far this has been an American story. Our own data adds a twist we did not expect to be this sharp.

Since March we have compared analyst targets against market prices daily on two markets — Helsinki and New York. In the spring measurement window (March 17 to June 9, roughly 1,400 company-days), US companies were promised a median 18 percent of upside. Finnish companies were promised two and a half percent.

The gap survives contact with what prices then did. The systematic error of US targets over the window was +16 percentage points — toward optimism, regardless of regime, and in falling markets slightly larger than in rising ones. The error of Finnish targets was +1 percentage point. In our spring data, a Helsinki price target was, for practical purposes, unbiased.

The international literature knows this pattern. Bradshaw, Huang and Tan (Journal of Accounting Research, 2019) compared price targets across 41 countries and found that the degree of optimism varies systematically from country to country, in ways connected to differences in market environments and institutions. Optimism, then, is not a trait of the analyst but a property of the environment. It carries a passport.

Honesty demands the same caveat as in our earlier essays: our two markets differ in more than geography. Half of the US panel is mega-cap technology; the Finnish side is mid-cap industrials and financials. Some of the gap may belong to sectors, not countries. But the direction matches the international evidence, and the gap is too wide to vanish entirely into composition: same instrument, two markets, fifteen percentage points of difference in honesty.

For the reader, the difference means that the same number means different things on different exchanges. In our spring data, a Helsinki price target described what followed roughly as written. A US target read systematically high — the literature and our own data agree on the size of the lean, roughly ten to fifteen percentage points.

A leaning instrument is still an instrument

It would be easy to stop here with a plain conclusion: price targets lean upward, so ignore them. The data says otherwise, and this is the essay's most important turn.

A biased instrument is not a broken instrument. A thermometer that always reads one degree high is perfectly usable — as long as you know to subtract the degree. And beneath the lean, price targets carry real information: in our spring data, US analyst targets were on the correct side of the subsequent price move on 61 percent of company-days. A coin manages fifty. The level is off, but the direction has signal.

This is where our five AI models enter — not as the hero, but as the counterweight. Our panel's raw DCF estimate leans the opposite way: over the measurement window its error was −8 percentage points, on the side of pessimism. Two instruments, two opposite leans.

And then something happens that we consider the single most interesting result of the spring's measurement work: combine the analyst view with the AI estimate — roughly three-quarters weight on the analyst, one quarter on the AI — and the blend's average error comes in below both of its ingredients. In the US, where analyst optimism is largest, the AI's weight rises and the benefit grows; in Finland, where analysts are already honest, there is little to correct and the AI's added value shrinks toward zero. The AI's value in this task was never that it forecasts better — alone, it loses to the analyst clearly. Its value is that it errs in the other direction.

When two instruments err in opposite directions and for different reasons, their errors partly cancel in combination — which is why the whole can be more accurate than either of its parts.

Back to Nvidia

Return to Thursday, July 9, and that $301.62.

Now we can read it. It is not a prediction anyone expects to come true — historically, two out of three such promises go unmet by their deadline. It is a favorable scenario, expressed through a convention that in the American environment leans a dozen percentage points upward, and that updates to follow the price at a fixed distance. Subtract all of that and something real remains: the view of 58 professionals that Nvidia's direction is more likely up than down. That is a genuine signal — just a far smaller one than the number 49 percent suggests.

You get the most out of a price target by looking at its level less often and its changes more often. The level carries the weight of the convention; the change carries the news — and it is around target revisions, research finds, that prices actually react. When a target falls while the price rises, the change carries information the level does not show. And when the spread across 58 analysts is wide, it tells you the same thing the spread across our five AI models has told us in earlier essays: the valuation is genuinely contested, and the average does not settle the dispute — it hides it.

The stock that is always on sale is not cheap. The discount lives in the instrument, not in the market. But an instrument whose lean you know beats no instrument at all — and between two instruments leaning in opposite directions, our spring data suggests, you see more clearly than from behind either one.

The measurements in this essay cover roughly 1,400 company-days from March 17 to June 9, 2026 (24 companies, Helsinki and New York). Target-price errors were measured against 30-day realized prices even though targets carry a 12-month horizon — the short window treats every series identically, but a full verdict requires the full horizon, which our data reaches next spring. Blend weights are partly fitted on the same data they are reported on; in a held-out test window the result held. Company-level figures in the opening reflect July 9, 2026.

AI Investor Barometer is an experimental research tool, not investment advice.