The rules
Methodology
These are the rules the system actually runs in production. They are read straight from its live configuration rather than retyped by hand, so this page cannot drift from reality.
Every step, in order
- 1Pull the full US listed universe (Nasdaq, NYSE, AMEX) — several thousand stocks
- 2Score every liquid stock on session signals: gap, relative volume, ATR, 48h news
- 3Lane A keeps the top 80 quantitative scores, lane B adds 20 on news and volume
- 4Perplexity researches the news, then GPT-4.1, Claude and Perplexity each vote
- 5Adjust for the market regime (SPY, VIX, breadth) and penalise overextended moves
- 6Publish up to 3 stocks from different sectors, each with its conviction level
- 7Measure open to close on the target session and compare to SPY on that same day
What makes up the final score
- Quantitative analysis (volume, gaps, trend)40%
- Three-AI consensus, weighted equally50%
- Quality of the identified catalysts10%
Penalties are then subtracted for excessive volatility, thin liquidity, strong disagreement between the models, incomplete data, imminent earnings, or a stock that has already run up a long way with no new information behind it.
Publication thresholds
- Minimum score to be published68
- Minimum confidence0.62
- Maximum stocks published per day3
- Finalists sent to the AIs each morning100
In a nervous market the minimum score is raised automatically and the liquidity requirement doubles. Falling short of a bar lowers the stated conviction, it does not hide the stock. We only withhold a pick entirely when too few models answered, the data was unreliable, the stock was too illiquid, or no model expected it to rise.
What the conviction label means
A stock that misses a threshold is still named, but it is named honestly. The label is fixed before the session opens and never changes afterwards, which is what makes it possible to check, on the numbers page, whether our confidence is worth anything.
- High conviction — Cleared every publication rule.
- Moderate conviction — Just short of the score or confidence bar.
- Low conviction — Well short of the bars. Published for continuity, not as a strong call.
How performance is measured
Every published stock is bought at the opening price of its target session and sold at the closing price of that same session, minus 0.1% in simulated trading costs. The result is compared to the S&P 500 measured exactly the same way, on the same day. The prices used are always the ones from that specific session, never the prices on the day the calculation happens to run.