AI Signals — Week 47, Nov 16–20, 2026
- Every single model turned more bullish this week — a rare unanimous drift that likely reflects falling market prices as much as genuine conviction shifts.
- Gemini leads the optimism league at +9.8% average upside, while GPT remains the panel's lone bear at -3.4%, a gap of 13 percentage points that should give any consensus-follower pause.
- NVDA carries the week's most fractured signal: a -9.4% downside consensus masks a dispersion score of 0.49 — the models are essentially arguing with each other in real time.
- TSLA and AAPL sit at -55.9% and -41.2% implied downside respectively, yet zero dispersion on TSLA suggests the models have reached an unusually cold consensus on a notoriously divisive name.
- DeepSeek costs $2.98 per thousand valuations versus Claude's $49.11 — a 16x price gap for a model running at near-zero upside bias, raising hard questions about what premium inference actually buys.
The Big Picture
The AI valuation panel covering 24 companies across Finnish and US markets delivered a quietly remarkable result this week: every model, without exception, shifted its aggregate bias upward. Claude moved from +4.9% to +7.1%, Gemini from +6.6% to +9.8%, GPT from -5.5% to -3.4%, DeepSeek from -1.1% to +0.6%, and Grok from -1.9% to -0.1%. A coordinated bullish drift across five architecturally distinct models sounds like a signal. It probably isn't — or at least, not purely.
Remember the data unit: bias_shift compares valuation gaps computed against each week's own market prices. If the market sold off this week, every model's implied upside mechanically widens even if the models revised nothing. The honest read is that this week's uniform bullish drift is a joint product of market price movement and model revision — and we cannot cleanly separate the two. What we can say is that no model moved against the tide, which is itself a behavioral observation worth filing.
The spread between the most bullish and most bearish model sits at 13.2 percentage points (Gemini at +9.8% versus GPT at -3.4%). That is a meaningful disagreement. When you pay for five models, you are partly paying for that spread — it is where the information lives.
Trends
The trends data this week is thin: only ORNBV (Orion Oyj B) registers a directional run, and it is a falling one — 3 down-days out of 4 with a range of 32.2% across the week. That range figure is the real story. A pharmaceutical name swinging 32 percentage points in consensus target price over four trading days is not a trend; it is a model in distress. The single up-day prevents this from being a clean bearish momentum signal, but the volatility suggests the models are struggling to anchor Orion's valuation — possibly because the stock's spot price of €100.00 with a consensus target of exactly €100.00 (upside: 0.0%, null dispersion) implies the models have effectively thrown up their hands and converged on a non-answer. A null dispersion score alongside a flat consensus is not agreement — it is the absence of information.
Sector Signals
The clearest rotation story this week is the collapse in healthcare, which shed 7.3 percentage points of implied upside to land at exactly 0.0% — the worst single-week deterioration across all sectors. With only two companies in the bucket, this is heavily influenced by Orion's implosion and JNJ's persistent -13.1% downside reading. Treat single- and dual-company sectors as individual stock signals wearing a sector costume.
Industrials gained the most ground, up +4.6pp to +19.4% implied upside — the highest absolute reading of any sector this week. Technology also firmed, rising +3.8pp to +14.6%, driven in part by NOKIA's +7.8% target price revision, the largest pure estimate change in the dataset this week. That Nokia revision is notable precisely because it is a model revision, not a price effect — the models genuinely upgraded their view of the network equipment maker.
Consumer continues to be the panel's least-loved sector at -26.0% implied downside, with AAPL at -41.2% and PG at -17.8% dragging the average into deeply negative territory. The models have been consistently bearish on consumer defensives for weeks; at some point, persistent model bearishness on a sector becomes a data point about model assumptions rather than sector fundamentals.
What the Models Reveal About Themselves
The cost-versus-behavior table this week is almost too clean. Claude, the most expensive model at $49.11 per 1,000 valuations, is also the most bullish at +7.1% average upside and carries the highest confidence score at 0.74. GPT, at $7.05/1K, is the most bearish at -3.4% and the least confident at 0.62. DeepSeek, at $2.98/1K, sits near-neutral at +0.6%.
One behavioral signature stands out: GPT's terminal growth dispersion is extraordinarily tight at a standard deviation of just 0.14 percentage points, and its CAGR dispersion (2.91pp) is the lowest on the panel. GPT is not just bearish — it is *consistently* bearish, applying near-identical discount frameworks across very different businesses. That is a sign of a model with strong priors that resist company-specific evidence. Gemini, by contrast, shows the widest CAGR dispersion at 6.3pp, suggesting it is genuinely differentiating between growth profiles rather than applying a house view.
Gemini's cap rate of 13.3% and Claude's 12.9% are the lowest on the panel, meaning these models are least likely to have their terminal values constrained by the sanity guard. GPT's 21.2% cap rate is the highest — one in five of its valuations is hitting the ceiling, which may be suppressing what would otherwise be even more bearish outputs.
Where the Framework Breaks
NVDA is the week's most instructive failure. The consensus target of $132.24 against a spot of $145.98 implies -9.4% downside — but the dispersion score of 0.492 means the models are almost maximally split. Someone on this panel thinks Nvidia is deeply undervalued; someone else thinks it is significantly overpriced. The consensus number is the average of two opposite views, which makes it nearly meaningless as a signal. This is precisely the case where publishing a single consensus figure does more harm than good. The right output for NVDA this week is not a number — it is a distribution, and a wide one.
At the other extreme, TSLA shows zero dispersion at -55.9% implied downside. Five models, architecturally distinct, all agree that Tesla is trading at roughly double its fair value. That is either a genuine insight or a shared blind spot baked into training data. The framework cannot tell you which.
The Model Scorecard
| Model | Avg Upside | Bias Shift | Cap Rate | Valid % | Cost/1K |
|---|---|---|---|---|---|
| gemini | +9.8% | +3.3pp | 13.3% | 94.2% | $26.31 |
| claude | +7.1% | +2.2pp | 12.9% | 96.7% | $49.11 |
| deepseek | +0.6% | +1.7pp | 15.3% | 98.3% | $2.98 |
| grok | -0.1% | +1.7pp | 16.9% | 98.3% | $8.47 |
| gpt | -3.4% | +2.2pp | 21.2% | 98.3% | $7.05 |