The premise is simple: markets move on news, but most news is noise. If a system could read every headline the moment it lands, throw away the 90% that does not matter, and act only on the signal that remains -- calmly, consistently, without the fear or greed that trips up humans -- what would it do?
Stock Intelligent is our answer. It is an automated market-analysis system that reads the news, weighs it, forms a conviction, and places simulated trades in a paper portfolio. This article walks through how it works, end to end, without the jargon: from the moment a story is published to the moment a position changes.
Today it trades on paper only. That is deliberate. We are proving the thinking is sound before a single real dollar is ever at stake.
The Big Picture
Stock Intelligent is a pipeline. News comes in one end, a trading decision comes out the other, and in between are a series of filters and judgements -- each one designed to be a little more selective than the last.
graph LR
A[News Sources] --> B[Noise Filter]
B --> C[LLM Analysis]
C --> D[Conviction Score]
D --> E[Decision Gates]
E --> F[Shadow Trading Engine]
F --> G[Paper Portfolio]
style A fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style B fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style C fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style D fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style E fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style F fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style G fill:#1f1f1f,stroke:#2a2a2a,color:#eee
Think of it as a funnel. Hundreds of articles enter. A handful of high-conviction signals come out. Everything in between is the system deciding what deserves its attention.
Step 1: Reading the News
The system continuously pulls company news for the stocks it follows. Our primary source is a financial news feed that provides fresh, ticker-tagged articles, and we enrich each story with a second data source that adds sentiment and context.
This is the raw material. On a busy day it can be dozens of articles per stock -- earnings coverage, analyst notes, product announcements, and a great deal of recycled commentary that says nothing new.
Step 2: Separating Signal From Noise
Most headlines are not worth trading on. Before a single article reaches the expensive analysis stage, it passes through a seven-part filter. Each gate asks a simple question, and an article has to pass all of them to survive.
graph TD
A[Incoming Article] --> G1{About this stock?}
G1 --> G2{Recent enough?}
G2 --> G3{Not seen before?}
G3 --> G4{Not a reworded copy?}
G4 --> G5{Relevant enough?}
G5 --> G6{Sentiment material?}
G6 --> G7{Real content?}
G7 -->|passes all| P[Send for analysis]
G1 -.fails any gate.-> X[Discard]
G2 -.-> X
G3 -.-> X
G4 -.-> X
G5 -.-> X
G6 -.-> X
G7 -.-> X
style A fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style P fill:#1f1f1f,stroke:#2a2a2a,color:#eee
style X fill:#3a1f1f,stroke:#5a2a2a,color:#eee
The filter checks whether the article is genuinely about the company, whether it is recent, whether we have already processed it, whether it is just a reworded copy of a story we have seen, whether it clears a relevance bar, whether its sentiment is material enough to matter, and whether it has real content rather than a headline stub.
This step is quietly one of the most important in the whole system. Analysis is not free -- so the discipline of throwing away weak articles early keeps the system both faster and sharper.
Step 3: The LLM Reads What Survives
The articles that make it through go to a large language model for genuine reading comprehension. This is where the system moves beyond keyword matching and starts to understand what a story actually means for the stock.
For each article, the model produces a structured read: is the sentiment positive or negative, how big is the likely price impact, and how confident is it? For the strongest or most ambiguous stories, the system escalates to a deeper, second-pass analysis rather than trusting a snap judgement. Weak or routine articles take the fast path. That way the expensive, careful thinking is spent only where it can change the outcome.
Crucially, the system also checks whether the news is already "priced in." A glowing story about a stock that has already run up 15% this month is far less tradeable than the same story about a stock that has not moved. Good news everyone already knows is not an edge.
Step 4: Forming a Conviction
Individual article reads are not decisions -- they are evidence. The conviction engine gathers all the recent, weighted signals for a stock and rolls them into a single number: the conviction score.
A strongly positive score leans bullish. A strongly negative score leans bearish. Scores near zero mean the news is mixed or unremarkable, and the right move is to do nothing. The score is deliberately built to be boring and repeatable: the same evidence produces the same number, every time.
| Conviction score | Interpretation | Action |
|---|---|---|
| High positive | Strong bullish case | Open or add to a position |
| Mildly negative | Softening case | Trim the position |
| Strongly negative | Bearish case | Exit the position fully |
| Near zero | Mixed or immaterial | Hold, do nothing |
Step 5: The Decision Gates
A conviction score is a strong opinion, but an opinion is not yet a trade. Before anything happens, the signal has to clear a series of safety gates -- the checks that separate a disciplined system from an impulsive one.
The system asks: Has the price already drifted too far from where the news landed (am I chasing)? Have I already acted on this same stock too recently (am I overtrading)? Do I have the cash reserve to support this? Is this signal genuinely new, or is it just an echo of a decision I already made in this same run?
That last question deserves a note, because getting it wrong was one of the system's early flaws. The engine used to accidentally count its own fresh signal as a reason to hold back -- effectively talking itself out of every trade. We fixed that so it now correctly ignores its own reflection and acts on the real history.
Only a signal that clears every gate becomes an order.
Step 6: The Shadow Trading Engine
When an order clears the gates, it reaches the shadow trading engine -- the part that actually manages the paper portfolio. "Shadow" means it trades with simulated money against live market prices. Every position, every gain, every loss is real in its arithmetic; only the dollars are imaginary.
The engine follows a few firm rules that reflect how a careful investor behaves:
- Entries are sized by conviction. A strong signal commits more; a moderate one commits less.
- Exits are proportional. A softening story trims a slice of the position; a clearly bearish one closes it entirely. It never sells more than it holds, and it never leaves an untradeable sliver behind.
- It never double-acts. If the same instruction somehow arrives twice, the engine recognises it and acts once. A trading system that occasionally trades twice is worse than useless.
- It cannot overspend. A purchase that would push cash below zero is refused, not fudged.
sequenceDiagram
participant N as News
participant A as Analysis + Conviction
participant G as Decision Gates
participant E as Shadow Engine
participant P as Paper Portfolio
N->>A: Filtered articles
A->>G: Conviction score
G->>E: Approved order
E->>E: Check: not a duplicate, funds available
E->>P: Update position + cash
E->>A: Confirm executed
Note over A,P: Every decision is logged, executed or not
Step 7: Remembering Everything
A system you cannot audit is a system you cannot trust. So Stock Intelligent writes down its reasoning for every stock it considers, on every run -- not just the trades it makes, but the trades it decides not to make, and why.
If it passed on a stock because the signal was too weak, that is recorded. If it held back because it had traded too recently, that is recorded too. Every executed trade and every paper position is tagged with the exact version of the strategy that produced it. When we later ask "why did the system do that in September?", we can answer precisely, because the system kept the receipts.
Why Paper Trading First
It would be easy -- and premature -- to point this system at a real brokerage account today. We are choosing not to.
The goal of the current phase is trust. We want to watch the system make hundreds of decisions, compare its predictions against what the market actually did, and satisfy ourselves that the thinking holds up before real money is involved. Paper trading gives us a faithful dress rehearsal: the same prices, the same logic, the same discipline -- with none of the downside if we are still wrong.
What We Learned
Selectivity beats volume. The single biggest lever was not smarter analysis -- it was throwing away more articles, earlier. A system that reads everything and acts on little needs a ruthless front-end filter.
The hard part is deciding when to do nothing. Opening a position is easy. Knowing when a signal is stale, already priced in, or simply an echo of a decision you have already made -- that is where a disciplined system earns its keep.
Auditability is a feature, not overhead. Recording why the system passed on a trade turned out to be as valuable as recording the trades themselves. It is how we tell "correctly cautious" apart from "quietly broken."
Safe failure beats clever failure. When a component could not confirm a trade was unique, we designed it to refuse rather than risk trading twice. In a trading system, the boring, conservative default is the right one.
What is Next
We are now collecting live shadow-trading results against a clean paper portfolio. Over the coming weeks we will measure how well the system's convictions line up with real market movements, tune the thresholds where the data tells us to, and build the track record that any move toward real capital would have to earn.
The thesis is simple, and it has not changed: read everything, act on almost nothing, and be able to explain every call. The market rewards patience and punishes noise. We built a system that is, by design, very good at ignoring things.
Follow the System's Thinking, Week by Week
The best way to judge a system like this is not to read about how it works once -- it is to watch it reason in public, over time. That is exactly what our market commentary feed, The Analyst, is for. Each week it publishes what the engine saw, what it concluded, and how those calls actually played out -- including the uncomfortable parts.
A recent entry is a good example of that candour: it documents the engine correctly holding maximum conviction into NVIDIA's earnings, while openly examining an "accuracy paradox" in the data -- strong multi-day directional calls paired with weaker next-day precision. We publish the questions we are still working through, not just the wins.
Subscribe to The Analyst: ecupse.com/blog/the-analyst
Follow along as the shadow portfolio builds its track record in the open.
This article was written by the Stock Intelligent team. The system it describes is live and processing real market news, trading a simulated portfolio while we build the evidence base for what comes next. Nothing here is financial advice.