AI Signals — Weekend Read: The gap that closed itself

Weekend Read #12 — August 2, 2026

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

On the morning of Friday, July 31, our pipeline finished its daily pass over twelve Finnish large caps and produced a number it had never produced before. The panel's median valuation gap for Finland — how far the models' own estimates sit from market prices — came out at +9.7 percent. The same morning, the median analyst consensus target for the same twelve companies implied +7.6 percent.

Two instruments built on entirely different foundations — one a panel of five language models feeding a deterministic DCF engine, the other hundreds of professional analysts with company access, sector experience and quarterly calls — were reading the Finnish market to within two percentage points of each other.

Then look west. For the twelve US companies in the same panel, the same models produced a median gap of −8.6 percent while analysts stood at +18.2 percent. Twenty-seven percentage points apart. Same models, same method, same morning.

For five months this project has measured how AI models value stocks. This week, for the first time, the interesting question stopped being why the AI is so bearish. It became: why do the two instruments agree in Helsinki and disagree in New York?

First, the confession

Before reading anything into that Finnish convergence, you deserve to know why it cannot be an artifact — because until three weeks ago, it could have been.

Until July 14, the headline estimate on this site was a blend: 70 percent model view, 30 percent analyst consensus. It was an honest, documented calibration choice — but it meant our number always carried a piece of the thing it was being compared against. If a blended estimate drifts toward analysts, nothing has been learned; the anchor is doing the drifting.

On July 14 we removed it. The published estimate is now the models' own value with no analyst input anywhere in the calculation — and at the same time we corrected three technical conventions (mid-year discounting, true 12-month targets rolled forward at the cost of equity, and a CAPM anchor that had been overcharging every company by a percentage point). Those corrections moved the level up mechanically, and we published the size of that mechanical shift before deploying it. Both series remain stored; the change is marked on every chart.

Which is exactly why this week's reading matters. When a pure model estimate — with the benchmark surgically removed from its bloodstream — lands within two points of the analyst consensus, the agreement is earned, not engineered.

Helsinki: two instruments, one reading

What does it mean when they agree?

Not that both are right. Two thermometers showing the same temperature can both be broken. But it narrows the possibilities in a useful way. The models reach their number through discounted cash flows built from growth, margin and discount-rate assumptions, anchored to reported fundamentals. Analysts reach theirs mostly through forward multiples, guidance and sector context. When two methods that share almost no machinery produce the same answer, the likeliest explanation is that the answer is not very sensitive to the method — that Finnish large caps are, at current prices, simply not contentious.

There is supporting evidence for trusting the Finnish reading in particular. In our spring measurement window, Finnish analyst targets erred toward optimism by roughly one percentage point — effectively unbiased, in a literature where US targets routinely run double digits hot. The Finnish consensus, in other words, has behaved like a calibrated instrument. Our panel now agrees with the calibrated instrument. That is about as close to external validation as a five-month-old measurement system can get.

It was not always so. In March, when this project started, the panel's Finnish gap sat around minus fourteen percent while analysts promised plus nine. Some of the closing was market prices rising toward estimates. Some was the July convention corrections lifting the level honestly. And some — the part we can see in the assumptions themselves — was the models' inputs settling as prompt and engine stopped nudging them downward. The gap closed from both ends.

New York: the 27-point disagreement

If Finland is settled, the United States is anything but — and the disagreement is not evenly spread. It concentrates precisely where the last two years of equity returns have concentrated: the mega caps. On Friday's numbers, the model consensus put Apple 52 percent below its market price and Tesla 48 percent below. The analyst consensus for the same two companies pointed firmly upward.

It is tempting to declare one side deluded, and readers of a certain temperament will happily pick which. The honest reading is more careful, and it cuts both ways.

The case for taking the models seriously: a discounted cash flow forces every valuation to answer one question — what future cash, discounted at what risk? For the largest US companies, current prices require growth and margin combinations at the edge of what the models' fundamentally anchored assumptions will produce, even after our July corrections removed the engine's conservative biases. Meanwhile, the same US analyst consensus erred sixteen percentage points toward optimism in our spring window — the documented, decades-old lean of the target-price convention, which last month's essay treated at length.

The case for taking the analysts seriously: DCF models are structurally cautious about businesses whose value sits in optionality — platforms, ecosystems, AI infrastructure — precisely the things the US mega caps monetize. A model that cannot imagine a new product category will always undervalue the company most likely to create one. The analysts' forward-multiple convention absorbs that optionality by pricing it the way the market currently does.

Note also what the panel does not do here: it does not speak with one voice. On Friday, GPT's median US-inclusive gap sat at −6.5 percent while DeepSeek's sat at +1.1 — the five models disagree with each other about levels even as they collectively disagree with the analyst consensus. The disagreement is layered, and each layer is measurable.

The check we already wrote

Here is the part we can promise rather than argue. In early July we preregistered a measurement: from September, once three months of clean data under the current engine exist, we score both instruments — the panel's pure estimates and the analyst consensus — against realized prices, by market, with the methodology fixed before the answer is known. The commitment is written down, the series are stored, and the result gets published whichever way it lands.

If US prices grind higher and the analyst targets prove closer, that is a finding about DCF's blind spot for optionality, and we will say so. If the mega caps revert toward the models' numbers, that is a finding about the target-price convention's lean, and we will say that too. A barometer does not win arguments; it makes them measurable.

Back to Friday morning

Two numbers, then, from the same morning run. In Helsinki, +9.7 against +7.6 — two unrelated instruments agreeing, for the first time, about an entire market. In New York, −8.6 against +18.2 — a 27-point disagreement sitting exactly on the companies that define this decade's equity story.

Five months ago this site's most common reader question was some version of why is the AI so negative about everything? That question has quietly expired. The panel is no longer negative about everything; it is negative about something specific — and it now agrees with professional consensus everywhere else. Whatever September's scoring says, the shape of the disagreement has become the product: not a machine that dislikes stocks, but a second opinion that concurs in one market and dissents, sharply and measurably, in the other.

The clearance sale, it turns out, was never storewide. It is one aisle. And both instruments are now pointing at it.

Model estimates are research output from an automated system, not investment advice. Methodology, engine changelog and the preregistered September measurement plan are documented on the methodology page.