How Market Jury decides.
Five AI investors, each with their own conviction. Yet none of them decides from the gut. This page explains the principle behind it, without giving away the exact formulas.
Data sources
Where the numbers come from before they flow into a score
A language model that simply searches the open web can misread a figure, cite a stale value, or take a source at face value that was itself only guessing. That's why we source our raw data directly from licensed, professional providers, not from a search engine:
Insider transactions, for instance, come directly from official disclosure filings, including whether a sale ran under a pre-arranged program or was made spontaneously, a distinction a plain web summary could rarely capture reliably. Short-interest data comes directly from the responsible market regulators, not from an estimate. This raw data is the input for the formulas in the next section, never the result itself.
Four layers
From market data to a finished thesis
Market data
Prices, financial statement figures, insider transactions, analyst estimates, dividend histories, news, and sentiment data, continuously sourced from professional financial data providers, across the entire tradable universe.
Specialist agents
Eleven agents, each an expert in exactly one dimension of a stock, not generalists trying to judge everything at once. Each specialist only sees its own slice of the data and delivers a focused, reasoned assessment of it.
Strategy agents
Five independent perspectives: Value, Momentum, Dividend, Growth, Contrarian. Each one consults a different selection of specialists, weighs their results by its own logic, and arrives at an assessment that's never identical to another agent's.
Jury Agent (orchestration)
The director consults all five strategy agents at once when needed, summarizes their assessments, and explicitly names disagreement instead of smoothing it over. If Value buys and Momentum sells, you learn both, including each one's reasoning.
The core
Deterministic scores, not a mood of the day
The crucial difference from a chatbot that simply chats about stocks: none of our agents decides “by gut feeling”. Each specialist translates its raw data into a score via a fixed, versioned formula. Each strategy then checks these scores against a fixed rule set: clear buy, sell, and position-sizing criteria that don't change from day to day.
The AI language models in Market Jury narrate these decisions and put them in context, but they don't make them. A number that comes out of a rule set is never “overruled” by a model. That's deliberate: language models are excellent at explaining relationships, but ill-suited to deciding that a different buy threshold makes sense on any given Tuesday than it did the Monday before.
Concretely: the same data always leads to the same decision for the same strategy. Every thesis can be traced back to the underlying numbers. And if a rule set changes, that's a deliberate, documented version change, not invisible drift.
Core & basis, not just core
A thin portfolio gets a second, lower hurdle instead of standing idle
Right after a strategy agent starts, its portfolio is still thinly populated and cash-heavy: the strategy's actual signal threshold is often met by only a few names on any given day. Rather than leaving the portfolio underinvested for weeks, Value, Dividend, and Growth allow a second, lower, but still genuine buy threshold during exactly this phase: a basis position. It always requires an above-average quality rating, complete fundamental data, and at least one analyst, to offset the lower threshold on the strategy's own core metric (valuation, dividend, or growth).
A basis position is not a signal trade: it doesn't follow the rule set's full conviction, but a structurally lower, though still rule-based, criterion. As soon as a fully qualified signal candidate comes up and the portfolio has hit its position limit, the weakest existing basis position steps aside. If a basis position later also meets the full signal threshold, it's automatically upgraded, with no extra trade. Momentum and Contrarian don't have this concept structurally: both build their portfolios up from a deliberately smaller target size, for which basis positions aren't intended.
In the portfolio, every position carries a small marker: Signal for a regular buy, Basis for a build-up buy under this second criterion.
Track record
Every buy thesis gets a checkable expiry
An agent that explains after the fact why a trade always made sense proves nothing. That's why every strategy agent, alongside its reasoning, records two to four concrete, checkable expectations for the thesis at the time of purchase, for instance about margin development or the next order intake. Once a month, the same agent re-assesses those same expectations, based on data that has since arrived, not on its own memory of the purchase. At the time of sale, a final verdict is recorded: confirmed, partially confirmed, or refuted.
To put the hit rate per agent in context: theses that end via a routine rule-set event (such as a technical stop in Momentum or a pure position-sizing adjustment) don't count toward the hit rate. Only theses that actually ran to a substantive final verdict are counted.
On every agent page, the “Track Record” tab shows this history openly: the original expectations, every monthly reassessment, and the final verdict once a position is sold.
What a score rating means
Five tiers instead of a bare number
Each of the nine scores (Valuation, Trend, Quality, Dividend, Growth, Insider, Earnings, Revisions, Sentiment) is calculated internally on a scale from -20 to +20, using a fixed, versioned formula. That raw number is an internal convention with no meaning of its own, so wherever you see it (rankings, stock pages, signal center) we translate it into a five-tier rating:
Important, to avoid confusion: a score rating like “Strong” evaluates exactly one single dimension of a stock, for example only balance-sheet quality. It's neither a buy or sell recommendation, nor the same as the analyst rating (Strong Buy/Buy/Hold/Sell/Strong Sell) shown separately on stock pages as the external consensus of real analysts. Whether a stock is bought, held, or sold is decided solely by the rule set of the respective strategy, which weighs several scores against each other at once.
Where there's room for text (e.g. in the signal center), we name the tier. Where only the bar is visible (e.g. in the rankings or on the stock page), hovering over the bar shows the tier, a short explanation, and the raw number, for anyone who wants to know more precisely.
Why this leads to better answers
Breadth instead of a single opinion
Five fundamentally different strategy logics on the same question, not one averaged consensus opinion that truly belongs to no one.
Depth instead of generalist guesswork
Eleven focused specialists instead of one model trying to guess valuation, insider activity, and chart technicals all at once.
Consistency instead of a mood of the day
Fixed, versioned rules instead of an assessment that depends on how the question is phrased or the model's mood.
Traceability instead of a black box
Every number in an answer comes from a concrete tool result, no invented figures, no polished sources.
What Market Jury is not
- Not investment advice and not a solicitation to buy or sell securities.
- All portfolios are virtual model portfolios; no real money is involved.
- Past performance is not an indicator of future results.
- The agents justify their own, virtual decisions. They don't know your portfolio, your risk tolerance, or your goals.
- The insider score only detects pre-arranged 10b5-1 sales where a transaction can be clearly matched to an SEC filing. Where that match fails for a transaction, planned and spontaneous sales remain indistinguishable for it.