Every equity on our forecasting page carries a classification: price-above-value, price-at-value, or price-below-value. But we never publish a target price. No "fair value is $X." No DCF models. No earnings multiples applied to projected numbers. This is intentional, and understanding why requires understanding what we actually do instead.
The Core Idea
Traditional valuation asks: what is this company worth? We ask a different question: does the current market price embed assumptions that the evidence supports?
This is not a semantic distinction. It changes the entire analytical framework. Instead of building a model that outputs a number, our AI ensemble identifies what the price appears to assume, then tests those assumptions probabilistically — treating the current market price as the prior belief and updating it against evidence.
How It Works: Four Stages
The thesis assessment sits at the end of a long pipeline. Each stage feeds the next.
Analysis Lenses Identify What the Price Embeds
Our multi-model AI committee runs multiple lenses (Gravy Gauge, Myth Meter, Moat Mapper, etc.) against SEC filings, transcripts, and thesis documents. Each lens produces signal assessments grounded in evidence — generated entirely by an ensemble of AI analyst personas, not human analysts. Critically, lenses like the Myth Meter explicitly identify what the current valuation appears to require — for example, 'the ~35x P/E requires ~13-15% annual EPS growth sustained over multiple years.'
Monitoring Triggers Become Forecast Markets
Each lens produces monitoring triggers — specific, observable events that would escalate or de-escalate a signal. We convert the highest-information-gain triggers into binary prediction markets with clear resolution criteria. For instance: 'Will Airbnb's FY2026 revenue guidance imply growth above 12%?' — because 12% is roughly what the current multiple requires.
9-Model Ensemble Estimates Probabilities
Each market gets 9 independent model runs (3 Opus, 3 Sonnet, 3 Haiku) grounded in a structured prediction context extracted from the analysis. The ensemble produces a median probability and a model agreement score. No single run drives the output — we want the central tendency across diverse reasoning approaches.
Thesis Assessment Synthesizes the Probabilities
A final synthesis model reads all market probabilities, their directional implications (escalate vs. de-escalate), model agreement, information gain, and the cross-lens analysis. The system asks: do these probabilities collectively support or undermine the assumptions the price embeds? The output is a computational classification, not a human opinion.
The Three Classifications
The prediction ensemble suggests that the assumptions embedded in the current price are too pessimistic. Key de-escalation scenarios have higher probability than the price implies, or key escalation scenarios have lower probability.
The prediction ensemble suggests that the assumptions embedded in the current price are broadly consistent with the probability-weighted evidence. Escalation and de-escalation scenarios are roughly in balance relative to what the price appears to expect.
The prediction ensemble suggests that the assumptions embedded in the current price are too optimistic. Key escalation scenarios carry meaningful probability, or key de-escalation scenarios that would validate the price have low probability.
A Concrete Example
Consider our recent Airbnb analysis. The Myth Meter lens identified that at ~$121 (~35x trailing P/E), the valuation requires approximately 13-15% annual EPS growth sustained over multiple years. This comes from the combination of 8-10% revenue growth, flat margins, and ~5% annual share count reduction from buybacks.
So we created a prediction market: Will Airbnb's FY2026 revenue guidance imply growth above 12%? The 9-model ensemble assigned this a 16% probability with 0.94 model agreement.
That single data point is already informative: the ensemble sees an 84% probability that guidance will confirm the growth deceleration narrative. But we don't stop there — we synthesize across all seven ABNB markets. The regulatory escalation markets (Barcelona enforcement at 40%, Fifth Circuit ruling at 30%) add probability-weighted downside risk. The competitive displacement market (Booking.com US share at 6%) removes a bear case concern.
The thesis assessment synthesizes these into: the price appears to embed a growth reacceleration narrative that the ensemble assigns only 16% probability, while regulatory escalation markers carry cumulative ~58% probability of at least one firing. Classification: price-above-value.
Why No Price Target?
Three reasons, in order of importance:
1. We don't have the machinery
Our pipeline produces probabilistic assessments of discrete events — not revenue projections, margin models, or discounted cash flows. A target price would require a fundamentally different analytical apparatus. We would rather do one thing well than two things poorly.
2. False precision destroys credibility
A target price implies a level of certainty about the future that prediction markets explicitly reject. When we say "16% probability of growth above 12%," that's a calibrated uncertainty statement. When someone says "fair value is $105," that's a point estimate masquerading as precision. Small changes in discount rate or terminal growth assumptions swing DCF outputs by 30-50%. The precision is illusory.
3. Direction is more defensible than magnitude
Saying "the price embeds assumptions the evidence undermines" is a statement about the relationship between expectations and probabilities. It's testable: when the prediction markets resolve, we can measure whether the ensemble was well-calibrated. A target price is not testable in the same way — stocks can trade above or below "fair value" for years, and the analyst can always claim the market hasn't "caught up yet."
What Traditional Analysis Does vs. What We Do
The Confidence Layer
Every thesis assessment carries a confidence level (LOW, MEDIUM, HIGH) that reflects how much weight to put on the classification. Confidence depends on:
- Model agreement — Do the 9 runs in each market agree, or are they spread out? High agreement across high-information-gain markets increases confidence.
- Data staleness — If the most recent earnings are imminent or the analysis is months old, confidence drops. Pre-earnings assessments are inherently provisional.
- Unresolved debates — If the analysis lenses surfaced genuinely unresolved questions (structural vs. cyclical growth, for instance), that uncertainty propagates into the thesis.
A classification of price-above-value at HIGH confidence means something very different from the same classification at LOW confidence. We always show both.
Accountability Through Resolution
The prediction markets that feed our thesis assessments have defined resolution dates and criteria. When they resolve, we score them with Brier scores — a proper scoring rule that penalizes both miscalibration and overconfidence.
This creates a feedback loop that target-price analysis lacks. If we assign 16% to revenue guidance above 12% and it comes in at 14%, that's a miss we can measure and learn from. Over time, our calibration page shows whether our probability assessments are well-calibrated — whether events we call "20% likely" actually happen about 20% of the time.
The Honest Limitations
This approach has real limitations we want to be transparent about:
- No magnitude — We can say the price appears above value but cannot say by how much. This limits the usefulness for sizing decisions.
- Assumption identification is itself a judgment call — When the system identifies "the price embeds 13-15% EPS growth," that's the AI ensemble's assessment. Others might disagree about what the market is pricing in.
- Market-specific coverage — The quality of the assessment depends on how well our prediction markets cover the key variables. If a critical factor wasn't converted into a market, it's not in the probability-weighted assessment.
- LLM-generated probabilities — Our ensemble uses AI models, not prediction markets with real capital at stake. The probabilities reflect model reasoning, not market prices. We are working to validate these against real-world resolution data.
Why We Think This Is Better
Not better in an absolute sense — better for what we're trying to do. Traditional valuation serves a purpose. But for educational analysis that prioritizes transparency, calibration, and intellectual honesty:
- Probabilistic language forces us to quantify uncertainty instead of hiding it inside a single number
- Testable predictions create accountability that target prices never face
- The pipeline is auditable — you can trace every thesis assessment back through the prediction markets, the analysis lenses, and the SEC filings that grounded them
- We optimize for being well-calibrated over time, not for being "right" on any single call
When we say price-above-value, we are saying: the assumptions the market appears to embed have low probability of playing out according to our ensemble. That's a specific, testable, falsifiable claim. We think that's more honest than a target price — and we'll find out over time whether it's more useful.
See it in action