The Big Picture

This was a week of coordinated retreat. Not panic — the models did not suddenly flip bearish en masse — but a quiet, synchronized pulling-back of optimism that is arguably more telling than any single dramatic call. Every model in the panel registered a negative bias shift. The aggregate direction is unambiguous: AI valuation engines looked at the same 23 companies they assessed last week and decided, collectively, that prices had gotten ahead of fundamentals.

The average upside across the panel ranges from -1.6% (DeepSeek) to +2.7% (Claude), a spread so narrow it barely constitutes disagreement. When five structurally different models converge this tightly on near-zero expected returns, that is not consensus — it is a signal. The models are telling us that, at current prices, the margin of safety has largely evaporated.

What makes this week's data particularly interesting is the uniformity of the pullback. This is not one model having a bad week. This is a system-wide recalibration.

Trends

Two names bucked the bearish tide with multi-day rising consensus: UPM-Kymmene and PG (Procter & Gamble).

UPM is the more intriguing case. Three up-days against one down, with a 9.93% intra-week range in consensus target — that is an unusually wide swing for a paper and pulp company. The models are not quietly drifting upward; they are actively reassessing. Given that UPM's consensus target still implies -17% downside to spot, this rising trend reflects models reducing their bearishness rather than turning bullish. The direction of travel matters, but the destination remains uncomfortable.

PG's rising trend is quieter — a 1.4% range suggests gentle, consistent upward nudging rather than conviction. For a defensive consumer staples name, this kind of slow drift is exactly what you would expect when models sense macro uncertainty and rotate toward quality. The problem is that PG's consensus still sits at -10% downside, so even the models' most optimistic lean cannot make the valuation work at current prices.

Sector Signals

The sector rotation table this week reads like a risk-off playbook with one conspicuous exception.

Energy is that exception. The sector improved by +9.2 points — by far the largest positive shift — moving from -35.8% to -26.6% implied upside. That sounds like a rehabilitation story until you remember that -26.6% is still catastrophically bearish. The XOM target price revision of +48.9% week-on-week is the mechanical driver here, and it demands scrutiny. A near-50% jump in a single week's consensus target for one of the world's most-analyzed companies is not a valuation insight — it is a model artifact. Something in the input data or prompt context shifted dramatically, and the output moved accordingly. This is the kind of number that should trigger a data audit, not a buy order.

Financials fell the hardest in relative terms, dropping -3.4 points to -2.1% implied upside. Last week financials were the one sector where models saw modest positive returns; this week that cushion is gone. With JPM sitting at -18% downside in consensus, the models are clearly uncomfortable with bank valuations at current levels — likely reflecting rate uncertainty and credit cycle concerns embedded in their DCF assumptions.

Technology remains the panel's favorite sector at +14.6% implied upside, though even here the trend is deteriorating (-2.8 points from last week). The sector's optimism is increasingly concentrated: TIETO, MSFT, META, and GOOGL all show positive upside, while AAPL at -14% downside and NOKIA at -22% downside drag the average. Technology is not a monolith — the models are making sharp distinctions within it.

Note that materials (single company: UPM) and telecom (single company: ELISA) each represent one name. Treat their sector-level numbers as individual stock signals, not genuine sector reads.

What the Models Reveal About Themselves

The bias shift data is this week's most important story. GPT moved from +6.9% to +1.3% — a -5.6 point collapse in a single week. That is not a gradual recalibration; that is a model that encountered something in this week's data and fundamentally repriced its worldview. Claude (-2.4) and Gemini (-2.6) moved in lockstep, suggesting a shared sensitivity to whatever macro or earnings signals entered the context window.

DeepSeek remains the structural outlier. It was already the most bearish model last week at -1.0% and has drifted further negative to -1.6%. More revealing: DeepSeek is the only model with zero terminal growth rate variance (stddev = 0.0), meaning it applies an identical 2.0% terminal growth assumption to every company in the universe. This is not sophisticated — it is a blunt instrument. Yet DeepSeek's cost efficiency ($2.22 per 1,000 valuations versus Claude's $40.85) means it will remain in the panel. The question is whether its structural pessimism is wisdom or rigidity.

Gemini's 82.6% validity rate deserves a harder look. Nearly one in five valuations fails to produce a parseable output. Gemini also shows the highest CAGR dispersion (8.48% stddev) — when it does produce a number, that number swings wildly. High variance combined with high failure rate is a difficult combination to trust in a production system.

Where the Framework Breaks

The XOM target price revision of +48.9% in a single week is where the framework breaks most visibly. ExxonMobil is not a volatile small-cap where a single analyst note can reprice the name. It is a $500bn+ company with decades of analyst coverage. When AI models collectively revise their consensus target by nearly half in seven days, the culprit is almost certainly input sensitivity — a change in the oil price assumption, a different earnings figure in the prompt, or a shift in how the models are weighting near-term versus terminal cash flows.

This is the fundamental epistemological problem with AI-generated DCF valuations: the models are exquisitely sensitive to their inputs in ways that human analysts, with their institutional memory and anchoring biases, are not. That sensitivity can be a feature — faster updating — or a bug — noise amplification. The XOM case this week looks like the latter. A +48.9% target revision with the stock still implying -22% downside is not a coherent signal. It is a reminder that these models are tools, not oracles.

The Model Scorecard