← Signals·

Findings

What the observatory's own data has shown so far — dated, one number each, with a link to the full analysis.
Nov 2026 · preregistered measurement, interim

Analyst targets came closer to the price 90 days later than the model estimates.

≈14 vs ≈32 ppaverage miss, analysts vs models

551 company-days, forecasts made 14 July – 14 August 2026. Both were off in a consistent direction: analysts too optimistic in both markets, the models too high in Finland and too low in the United States. The tests fixed in advance did not settle either way; 23 of 64 forecast dates have matured.

Read: Readout →

Sep 2026

A model can change overnight without its name changing.

−7.4 → −0.5 ppDeepSeek's gap relative to the other four models, before and after 31 July 2026

On 31 July the provider began serving a reasoning variant under the same model name, and the panel's steadiest pessimist moved to the median. Switching reasoning off on 14 September cut the run time but did not bring the old level back. Such breaks are dated in the changelog.

Read: Weekend Read →

Aug 2026

The five models move as one.

≈1.1effective number of independent opinions, out of five

Day-to-day changes in the models' valuation gaps correlate at 0.91. A controlled experiment of 1,440 model calls with three prompt interventions moved the figure only from 1.05 to 1.07, and a frontier reasoning model run through the same protocol joined the group on arrival.

Read: Weekend Read →

Aug 2026

But their levels differ, persistently.

≈8 ppbetween the lowest and the highest model's average valuation gap

In August GPT's average gap to market prices was −6.5 % and DeepSeek's +1.1 %. The models move together, but each sits at its own level — the information is in the levels, not in five separate movements.

Read: Weekend Read →

Jul 2026

Analyst targets lean optimistic — mostly in the US.

+16 vs +1 ppanalyst target error toward optimism, US vs Finland

Measured over about 1,400 company-days from 17 March to 9 June 2026, against what prices then did. The 90-day measurement points the same way so far: analyst targets were too high by about 11 points in the US and 5 in Finland.

Read: Weekend Read →

Jun 2026

Similar numbers, very different reasoning.

368 vs 35words of reasoning per stock, Claude vs GPT

Across 60 outputs on one prompt and a scan of 8,257 valid outputs, the estimates correlated at about 90 % while the written reasoning did not. Self-reported confidence ran opposite to the detail of the reasoning: the most confident model wrote the least specific text.

Read: Weekend Read →

Findings describe how the models behave on this valuation task, measured on this site's own data; they are not a view on any stock. Each one comes from the analysis it links to. Engine and model changes after a finding are dated in the changelog. Not investment advice. Changelog

Findings — what five AI models showed on a valuation task — AI Investor Barometer