Transparent AI signal scoring.
MarketPicks.ai does not predict guaranteed returns. It produces research signals by combining weighted market inputs, then labels the result Bullish, Neutral, or Bearish with a confidence score.
This page documents the whole path from raw feed to visible label: which inputs are read, how they are weighted, how a rating threshold is decided, where the data comes from and how often it is refreshed, and what the model deliberately does not do. Every number shown on a signal card traces back to one of the inputs described below.
The four weighted inputs
News sentiment
Headline tone, recent news flow, and source-backed market catalysts.
Price momentum
Daily movement, weekly direction, and price pressure relative to recent context.
Trend velocity
Acceleration in discussion, unusual activity, and market attention.
Volatility risk adjustment
Risk, drawdown sensitivity, uncertainty, and event-driven instability.
Those four weights are the legacy weighted model: news sentiment at 35%, price momentum at 25%, trend velocity at 20%, and a volatility risk adjustment at 20%, which together produce a 0–100 confidence figure that is then mapped to Bullish at 58 or above, Bearish at 42 or below, and Neutral in between. It is kept as a readable backstop alongside the ensemble so the arithmetic behind a score can always be checked by hand.
The feature set the ensemble votes on
Instead of collapsing everything into one formula, the engine runs sixteen small decision trees. Each tree sees only a subset of features — the way a forest learner samples features — casts one bull, bear, or neutral vote with a weight, and writes down the reason for that vote. Those reasons are what appear as top reasons and contradictions on a signal page. The trees read the nine features below directly, and several of them read derived views built from them: sentiment divergence (headline tone against the weekly tape), trend structure (distance from the 50-day average confirmed by momentum), and a composite oscillator averaging momentum with trend velocity.
News sentiment
0–100 average of the tone of the recent headlines found for the symbol. No headlines means the score stays mid-range rather than being guessed.
Momentum
0–100 blend of day-over-day change and distance from the 50-day simple moving average.
Trend velocity
0–100 rate of change of the move, so a fading uptrend scores lower even while the price is still rising.
Volatility risk
0–100 measure of how wide the recent path has been. It enters the score as an adjustment, not a direction.
Unusual volume
0–100 comparison of current volume against the 20-day average — the footprint of attention arriving or leaving.
Range position
0–100 for where price sits inside its 52-week range, which separates a breakout attempt from a mid-range drift.
Day change
Raw percentage move since the previous close, used as confirmation or contradiction against the scored inputs.
Week change
Raw percentage move over the week, which is what the sentiment-divergence and follow-through votes compare headline tone against.
Distance from the 50-day average
Raw percentage above or below the 50-day SMA, used by the trend-structure and moving-average votes.
From votes to a label
Each vote contributes positively, negatively, or not at all according to its direction, weighted by how strongly that tree leaned. The weighted result is normalised to a 0–100 confidence starting from 50 as the undecided midpoint. A rating only becomes directional when one side has both the majority of votes and enough blended confidence to clear its threshold:
Bullish
More trees vote bull than bear and blended confidence reaches 56 or higher.
Bearish
More trees vote bear than bull and blended confidence lands at 44 or below.
Neutral
Every other case — including splits where neither side has both the votes and the confidence to justify a directional label.
Confidence labels
Confidence within 2 points of 50 reads as no edge, within 6 points as a small edge, within 12 as a real edge, and beyond that as a strong edge. The label describes distance from an undecided midpoint, not a win rate.
Data sources and refresh
Signals are computed from cached market data stored with the site: a quote table for price, change and volume, and a stored headline set for news tone. Scheduled jobs refresh those tables through the day, and the health page reports each feed's last successful run so staleness is visible rather than assumed away. Quote and news feeds count as healthy for 90 minutes after a run, degraded up to two hours, and down after that; the slower calendars run daily.
When a symbol has no cached row, the engine will attempt a single upstream quote lookup rather than render an empty card, and only then fall back to an explicit unavailable state. Signal results are cached for about a minute before being recomputed, so a reload picks up fresher inputs without hammering the feed.
The stock screener uses a separate snapshot: a full US-listed universe regenerated on a 30-minute schedule and filtered in the browser, which is why the screener page shows a timestamp for when prices were written rather than claiming a live feed.
What the methodology does not do
It does not place orders or know your account
MarketPicks has no access to positions, cost basis, tax situation, or time horizon. Nothing produced here can account for those, which is why nothing here is advice.
It does not forecast a return
A score and a confidence label are model output about current conditions. They are not a probability of profit, a price target, or a promise about what happens next.
It cannot score what it cannot see
When chart history for a symbol is too thin to compute moving averages, range, and volume properly, the engine runs in quote-only mode, limits what it claims technically, and caps confidence instead of reporting a full-strength number it did not earn.
It does not invent missing data
With no cached or fetchable quote for a symbol, the page returns an explicit empty state rather than a placeholder price. An honest gap is preferable to a fabricated one.
Headline tone is not fundamentals
News sentiment is read from headline text on a short window. It can be swayed by repetition, paywalls, or a slow news day, and it says nothing about balance sheets or valuation.
Scores go stale
Quote and news inputs come from scheduled refreshes. If a feed falls behind, the label on the page reflects the last data the engine actually received — the site's health checks treat stale feeds as degraded rather than pretending they are live.
How to use a signal
Read the label, then read the reasons. A Bullish rating with contradictions listed underneath it is a different trade of information than a Bullish rating where every tree agrees — the signal page shows both for that reason. Check the data-input panel as well: it states which feed produced the number and whether the engine was working from full chart history or a quote-only fallback, which directly affects how much weight the confidence figure deserves.
Signals are inputs to your own research, not conclusions. Where a score and your reading of the business disagree, the disagreement is the useful part — it is either a gap in the model's inputs or a gap in your thesis, and only further work tells you which.
Not financial advice. MarketPicks.ai is an impersonal research tool. AI output can be wrong. Investing involves risk of loss. Do your own research and consult a licensed professional. Full disclaimer.