AI Signals — Week 44, Oct 26–30, 2026
- Gemini completed only 39% of its valuations this week — a reliability collapse that makes its readings statistically unreliable for any serious cross-model comparison.
- GPT remains the most bearish model at -7.9% average upside, yet shifted the most bullish of any model this week (+4.0pp), suggesting its pessimism is slowly eroding under earnings-season pressure.
- The models collectively see AAPL as 55% overvalued and TSLA as 59% overvalued — yet their target prices fell further this week, deepening the disconnect with market prices.
- Telecom surged +10.4pp in model favor, driven entirely by a single company (Elisa), which means the 'sector signal' is really a single-stock signal wearing a sector costume.
The Big Picture
The AI model ensemble is sending a clear, uncomfortable message to equity bulls this week: the market has run ahead of fundamentals, and the gap is widest at the very top of the cap table. Across 24 companies and five models, the aggregate picture is one of modest-to-negative conviction. Three of the five models carry a negative average upside bias — GPT at -7.9%, Grok at -2.1%, and DeepSeek at -1.4% — while Claude and Gemini eke out thin positive readings of +2.7% and +3.2% respectively. The ensemble is not screaming crash, but it is not endorsing the current price level either.
The most striking single data point this week is not a model bias figure. It is the consensus target on AAPL: $156.35 against a spot price of $348.91, implying -55.2% downside. For TSLA, the gap is -59.3%, with a consensus target of $164.27 against $403.30. These are not rounding errors or model glitches — both show near-zero dispersion, meaning the models are in rare agreement. What this tells us is not necessarily that Apple and Tesla are about to halve, but that the models' DCF frameworks, anchored to earnings power and discount rates, simply cannot justify prices that embed years of speculative optionality. The models are not wrong to flag this; they are just speaking a different language than the market.
Trends
Four stocks showed directional momentum in model consensus across the week, and the pairings are instructive. On the rising side, ORNBV (Orion Oyj B) climbed on 3 of 4 days with a 7.5% intra-week range, and its target price was revised up +15.0% — one of the largest pure estimate revisions in the dataset. Wärtsilä (WRT1V) was even more consistent, rising on 3 of 4 days with zero down-days and a +4.5% target revision. These are not momentum trades; they represent genuine model re-rating, likely driven by updated earnings assumptions feeding through the DCF machinery.
On the falling side, META dropped on 3 of 4 days with an 8.33% range — the widest of any trending stock. This is notable because Meta still carries a +14.6% consensus upside, meaning the models are not bearish on it in absolute terms; they are becoming *less* bullish. The direction of travel matters as much as the level. AAPL also fell on 3 of 4 days, but with a narrow 1.77% range — a slow bleed rather than a repricing event, consistent with the -9.7% target price cut it received this week.
Sector Signals
The sector rotation table this week has a structural caveat worth stating plainly: several 'sectors' contain only one company. The +10.4pp surge in telecom is entirely attributable to Elisa (ELISA), whose consensus upside moved from 32.5% to 42.9%. Treat this as a stock call, not a sector thesis.
Healthcare's +6.8pp improvement is more meaningful, spanning two companies (JNJ and ORNBV), and JNJ received a +18.8% target price revision — the largest in the dataset this week. That is a substantial re-rating, and combined with Orion's upgrade, it suggests the models are finding more value in defensive healthcare names as growth multiples elsewhere look stretched.
Financials slipped -4.2pp, the sharpest sector deterioration of the week. With four companies in the bucket — including JPM at -12.3% consensus downside — the models are consistently skeptical of bank valuations at current prices. This is a recurring theme: financial models built on normalized earnings struggle to justify prices that embed rate-cycle optimism.
Technology, the largest sector at 8 companies, barely moved: +0.8pp to a near-flat +0.6% average upside. Beneath that calm surface lies enormous dispersion — GOOGL at +42.8% and AAPL at -55.2% sitting in the same bucket. The sector average is meaningless here; the distribution is what matters.
What the Models Reveal About Themselves
The most important behavioral story this week is GPT's +4.0pp bias shift — the largest single-model move in either direction. GPT moved from -11.9% to -7.9%, still the most bearish model but meaningfully less so. This shift reflects both market price movements and genuine estimate revisions, so it cannot be attributed purely to a change of analytical view. But the direction is consistent with a pattern we have observed before: GPT's deep pessimism tends to moderate during earnings seasons when hard numbers displace narrative.
DeepSeek moved the other way, dropping -1.7pp to -1.4%. This is worth watching. Since the provider-side switch to reasoning mode on 2026-07-31 moved DeepSeek's bias toward zero, any renewed drift into negative territory could signal the reasoning variant is re-calibrating toward more conservative assumptions. One week is not a trend, but it is a flag.
Claude ticked up +0.9pp to +2.7%, maintaining its position as the most consistently optimistic model with the highest confidence score (0.73). It is also the most expensive by a wide margin — $50.26 per 1,000 valuations versus GPT's $7.04 — and the slowest, averaging 39.7 seconds per valuation. Whether that premium buys genuine analytical edge or just more elaborate reasoning chains is the central question this platform exists to answer.
Gemini's reliability problem deserves its own paragraph. A 39.2% valid rate — meaning 60.8% of its attempts produced no usable output — is not a minor data quality issue. It is a fundamental question about whether Gemini belongs in a production ensemble at all. Its average confidence of 0.31 (versus Claude's 0.73) compounds the concern. When Gemini does produce a number, how much weight should it carry?
Where the Framework Breaks
JNJ is this week's most instructive failure case. Its consensus target price equals its spot price exactly — $275.20 — implying precisely 0.0% upside. The dispersion field is null, meaning the models produced outputs too inconsistent to calculate a spread. Then the target price was revised up +18.8% within the same week. What this sequence reveals is a model ensemble that is genuinely uncertain about JNJ's fair value, oscillating rather than converging. Johnson & Johnson is a complex business — litigation overhangs, MedTech cyclicality, pharma pipeline uncertainty — and the models are apparently unable to anchor on a stable earnings trajectory. The +18.8% revision is not a confident upgrade; it is noise masquerading as signal.
The Model Scorecard
| Model | Avg Upside | Bias Shift | Cap Rate | Valid % | Cost/1K |
|---|---|---|---|---|---|
| claude | +2.7% | +0.9pp | 12.7% | 98.3% | $50.26 |
| deepseek | -1.4% | -1.7pp | 16.8% | 99.2% | $2.90 |
| gemini | +3.2% | -0.4pp | 12.8% | 39.2% | $10.28 |
| gpt | -7.9% | +4.0pp | 19.3% | 99.2% | $7.04 |
| grok | -2.1% | +1.3pp | 12.6% | 99.2% | $8.46 |