AI Signals — Weekend Read: The machines were the bears

Weekend Read #10 — June 28, 2026

AI Investor Barometer runs five large language models over 23 listed companies every trading day. Each produces the assumptions behind a discounted-cash-flow valuation; a deterministic engine turns those into a model estimate. This week we set those estimates against the people whose job this has always been.

For 23 companies, every trading day, five AIs publish a valuation. And for the same 23 companies, dozens of human analysts publish theirs — the consensus price target you see quoted on every finance page. Two panels, same stocks, same dates. We had never put them side by side. Over roughly 1,400 company-days, with a realised price thirty days later, we finally did.

The cliché says the machines are reckless and the humans are prudent. The data says the opposite.

Three things to hold first, because they decide how much weight any of this carries. The realised window is thirty days against twelve-month targets — a glance, not a verdict. The entire sample sits inside a rising market, which, as we'll see, quietly flatters one side. And where we combine the two panels later, the weights are fitted in hindsight on this same data — illustrative, not a recipe. This is a research observatory, not investment advice.

Both are coin flips

Start with the unglamorous finding, because it frames everything else. Over the sample, the AI consensus called the direction of the next thirty days correctly 54.6% of the time. The analyst consensus: 54.4%. A statistical dead heat.

Neither panel can tell you which way a stock goes next month. That is not a knock on either — it is the well-worn result of half a century of forecasting research, which finds human target prices correct on direction barely more than half the time. Our analysts replicate it almost exactly. The AIs match them.

So if direction is a coin flip for both, the interesting question is not who is right. It is how each one is wrong.

The mirror

Here the two panels separate cleanly — and in the direction nobody expects.

The average AI valuation sat 2.1% below the market price. The average analyst target sat 10.8% above it. The stock actually moved +2.4%.

The machines were the pessimists. The humans were the optimists. The truth landed between them — closer to neither.

That analyst optimism is not a quirk of our sample. Sell-side target prices have carried a documented upward bias of roughly +9% for decades; ours come in at +10.8%, almost on the nose. What is new is the other side of the mirror: the AI panel is biased the opposite way. Where humans lean hopeful, the models lean cautious. Five language models, handed the same financial statements the analysts had, arrived systematically under the market rather than over it.

This is the single most distinctive thing we have found in four months of running the observatory. The AIs are not a faster analyst. They are an analyst with the sign of the bias flipped.

But the optimists won this round

Honesty requires the next paragraph, and it cuts against the AIs.

In this sample, the optimists were closer to the truth. On 57.4% of company-days the analyst target ended up nearer the realised price than the AI estimate did. The analysts' average miss was 14.5%; the AIs' was 19.0%. The humans were less wrong on magnitude.

Why? Because the market rose. When prices drift upward, an optimistic forecast is bailed out by the tide, and a cautious one is left behind. The AIs' pessimism was a liability precisely because nothing went wrong during the window. Our entire dataset is a bull market; in a drawdown, the sign on this paragraph flips. The fair reading is not "analysts are better." It is "in a rising market, optimism is rewarded, and the analysts supplied it."

So we have two panels that cannot call direction, biased in opposite directions, each better than the other under different skies. That sounds like a stalemate. It isn't.

Better together

The most useful result in the whole exercise is what happens when you stop treating the two as rivals and combine them.

A simple blend — roughly one-third AI estimate, two-thirds analyst target — beats both panels on their own:

The blend is more accurate than either ingredient and calls direction better than either. In forecasting terms, this is the cleanest possible signal that the AI carries information the analysts do not — and the analysts carry information the AI does not. If the models were just a noisier copy of the analyst view, mixing them in could only hurt. Instead it helps. The pessimism is not error; it is the missing correction to human optimism.

That, finally, is the role these models seem built for. Not to replace the analyst. To de-bias him.

Where it pays off

The blend also tells you when the AI is worth listening to, and the answer is precise: exactly as much as the humans are over-excited.

In the United States, where analyst targets ran a remarkable +18.7% above the market, the optimal mix leaned far more on the AI — and the blended miss fell to 12.1% with direction up at 62%, comfortably ahead of either panel alone. In Finland, where analysts were nearly sober (+2.2% above market), the AI added almost nothing; there was little optimism to correct.

The model's value, in other words, scales with the bias it is correcting. The AIs earn their keep loudest in the market where humans are most hopeful.

What this means

It would be easy, and wrong, to turn this into "AI beats Wall Street." It doesn't. On the only test the market actually pays for — calling direction — the machines and the humans are indistinguishable, and both are close to chance. On magnitude, in this rising market, the humans were closer.

The real finding is quieter and more durable. The five models are a structural counterweight to the optimism that has defined sell-side research for as long as there has been sell-side research. They are wrong in the opposite direction, and being wrong in the opposite direction turns out to be useful: combined with the analyst view, they sand off its hopeful edge and produce something better than either alone.

What we cannot yet tell you is how this survives a falling market — the one regime our four months have not contained. If the pattern holds, the AIs' caution should pay its largest dividend exactly when the analysts' optimism costs the most. That is the test we are waiting for, and the accuracy page is where it will show up first when the three-month window opens in July.

For now, the cliché can be retired. The reckless ones, this time, were the people. The machines were the bears.

*The figures above cover roughly 1,400 company-days with a 30-day realised horizon, drawn from the production database. The blend weights are fitted in-sample and are illustrative, not a strategy. Methodology: /methodology.*

— AI Investor Barometer