No black box and no hand-waving. Here’s the exact pipeline behind every analysis, the live data it runs on, how the ratings and targets are formed, and, just as importantly, where the limits are.
Not one generic prompt, five focused experts, then a manager that decides.
Fundamentals
Reads valuation (P/E vs sector), growth, margins and balance-sheet health to judge whether the business justifies the price.
Technical
Maps trend, momentum and key support/resistance from live price data, and locates the stock within its 52-week range.
Sentiment
Weighs recent company news, the narrative and the Wall Street analyst consensus to gauge how the market is leaning.
Risk
Stress-tests the downside, volatility, beta, event risk (earnings) and the specific things that could break the thesis.
Portfolio Manager
Synthesizes the four specialists into one decision: a rating, conviction score, entry zone, stop, targets and an options play.
A ranking model, graded against SPY.
What it publishes
Every market morning the model ranks 330+ large US stocks against each other from the previous close and posts the top 15 as buys and the bottom 15 as avoids. A buy is a hit if the stock beats SPY over the next 5 trading sessions; an avoid is a hit if it lags SPY. A tie is a miss. Returns include dividends. Each call is signed when it posts (ticker, start price, SPY’s close, resolve date), so it can be checked later for edits.
Compared with its peers, not the market
Scores are compared inside each sector and stripped of any tilt toward high- or low-beta stocks, so the board is not a disguised bet on tech, or on the market going up. It is a bet on which stocks beat comparable ones.
How it was tested
Walk-forward over 2010–2026 on daily prices and point-in-time SEC filings: each figure counts only from the day after it was filed, and each year is scored by rules fixed before that year. At the 5-day horizon its buys and avoids were right against SPY 52.2% of the time (buys 52.6%, avoids 51.8%), with a rank-correlation t-stat of 6.63 on non-overlapping windows. That is a small edge: most single calls will look like coin flips, and it only shows across hundreds of them.
Earnings dates without a calendar
The filings don’t say when results were announced, so the release is taken as the day trading volume spikes between a quarter’s end and its 10-Q or 10-K. Before that filing is in, a spike at least three times normal after the quarter should have ended counts, so a fresh report is read the day it happens.
What we can’t claim yet
Our own pre-set bar did not pass: a plain 12-month-momentum ranking scored about as well on average, though far less steadily, so the blend has not beaten momentum on every measure. This version was chosen by a fixed rule from the 2010–2023 results alone, among candidates written down before any was run, but only after an earlier version had been built and its results seen. The 5-day horizon was chosen after results at 30 days had been seen, and the 2024–2026 years (52.5% right) had already been looked at. Its signals come from published research, which tends to weaken once published. The universe is today’s list of companies, which flatters history. So the backtest is the design target, not proof. The proof is the forward record, which starts with the first live board.
The original net
Before this model, the board was made by the BrainStock neural net and a call was graded on whether the stock closed at or above its start price. Its calls stay on the public record under their own rule, counted separately, never blended with this one. The net still powers the on-demand forecasts on BrainStock.
The 7 signals, equally weighted
Each is computed only from prices, volume and SEC filings available by the close it ranks on. A stock missing a figure ranks as average on it.
The data behind every call
Quotes, fundamentals, analyst ratings and company news come from Finnhub in real time, the same class of feeds professional desks use. Charts are live. Nothing on the analyzer is mocked or hard-coded. When data for a field isn’t available, we show that rather than inventing a number.
How ratings & targets are formed
Each agent scores its domain, then the Portfolio Manager weighs them into a single Buy/Hold/Sell rating and a conviction score. Price targets are framed as scenarios, a bear, base and bull 12-month case, because a single number hides the risk. Conviction is shown honestly: a 55% call looks different from a 90% one.
How the options figures work
The AI suggests a structure (a call or put), an approximate strike, breakeven and max loss, and draws the payoff at expiry. When a live options data provider is connected, we then pull the real contract nearest that strike from the live options chain, actual mid premium, implied volatility, delta/theta, open interest and breakeven (quotes may be delayed). If live chain data isn’t available for a name, we fall back to a realized-volatility estimate and label it clearly as an estimate.
We publish our record
Every analysis is logged the moment it’s generated, entry price and timestamp, and scored against live prices afterward. We don’t delete the losers. You can audit the AI’s hit rate and average return yourself on the public track record.
What this is, and isn’t
YN Finance is an education and research tool, not a broker or a registered advisor, and nothing here is financial advice. The AI can be wrong, markets are uncertain, and past performance never guarantees future results. Use it to think faster and more thoroughly, then make your own decision.
See it on a real stock.
Run any ticker, then check the call against our public track record.