Methodology / Attention
How Attention finds your next research question
Attention brings recorded portfolio, watchlist and market signals into one review queue. Start with the reason on each card, check its source and date, then decide whether it deserves further research.
Open your attention queue →- Attention score
- How strongly the recorded signals contribute on a 0–100 scale.
- Queue rank
- How an item compares with other candidates, followed by a selection that favors variety.
- Reason severity
- The importance assigned to an individual reason. It is separate from the overall score.
Scoring Model
The shared scorer adjusts each signal for its type, confidence, severity and age. It combines positive contributions with diminishing additions: more evidence can raise the score, up to 100. A single strong signal can reach the top tier.
Signal A
60
Signal B
40
Combined
76
60 uses 60 points of the scale. The next contribution adds 40% of the remaining 40 points: another 16. The result is 76, not a 76% probability of loss. These are example inputs, not your holdings.
Show the calculation and timing rules
score = 100 × (1 − ∏(1 − min(1, contribution / 100)))
Only finite, positive contributions enter this combination. A matching emitted signal replaces its duplicate base input; signals of the same type within a 24-hour window are suppressed by the shared scorer.
For dated signals, the age adjustment has a three-day grace period, then multiplies strength by 0.9 for each additional day. An undated signal receives full weight. This is a calculation rule, not an expiry policy for every stored card.
Event-driven producers can supply their own scores. The queue preserves those contributions alongside the scheduled builder. A card’s creation time does not establish the age of all its underlying evidence.
- Watch
- 0 to below 50
- Warning
- 50 to below 75
- Critical
- 75 to 100
These labels describe recorded signal strength. A low score can reflect missing inputs; it does not establish that a holding is safe or that no action is needed.
Why a lower score can appear first
Queue rank compares available signals across the candidate set and reduces the influence of redundant, correlated signals. The final selection also favors different kinds of issues, so several similar high-score cards do not crowd out everything else.
Changing filters changes the comparison set. When there are too few comparable signals, ranking falls back to the attention score divided by 100. A ticker-filtered request uses score ordering directly. Queue rank is not a forecast of investment returns.
Active items respect your snooze and dismissal choices and any selected filters. Scores below 50 can appear. The page loads a bounded set of candidates, so it should not be read as an exhaustive inventory of every possible risk.
Data Inputs
Depending on the holding and available sources, reasons can include position concentration, factor exposure or drift, institutional ownership changes, insider activity, filing red flags, debt or litigation risk, and alerts. Each card shows the reasons actually recorded for it.
Company-specific signals do not apply uniformly to funds and cash. Regulatory filings describe earlier activity, and a successful brokerage sync does not prove that every research source is current. Scheduled refreshes and event-driven updates depend on their upstream data.
Explore alternatives with context
For supported negative signals, alternatives look for same-sector securities with a factor similarity score of at least 60 and at least three shared factors. The comparison then checks the selected signal. Pro shows up to one candidate; Premium shows up to three with comparison details.
The lower-confidence toggle can include candidates whose signal source is missing. Read that distinction before comparing them. A search can also return no matches.
Similarity does not guarantee factor neutrality, suitability or wash-sale safety. This lookup does not perform those checks and is not an instruction to replace a holding.
Limitations & Known Biases
- Missing or delayed inputs can understate the need for review. No card is not the same as no risk.
- Related signals may describe the same underlying event. Duplicate suppression and ranking adjustments do not establish statistical independence.
- The score is a prioritization heuristic. It has not been established here as a calibrated loss probability or a validated predictor of excess returns.
Verification and evidence
This explanation follows the scoring, selection and alternatives implementations. Tests check bounded combination, tier boundaries, missing inputs and presentation. Those checks establish software behavior; they do not establish investment performance or complete data coverage.
Citations & Research Basis
This page documents an implementation, rather than claiming an independently validated investment model. The following modules define the current behavior:
src/lib/analytics/attention.ts— contribution combination and age adjustment.src/services/attention-selection.tsandsrc/lib/research/rank_ensemble.ts— relative ranking and diversity.src/services/attention.ts— queue state, filters and score-order fallback.src/app/api/jobs/attention-queue/route.ts— scheduled builder;src/components/attention/attention-queue-list.tsx— presentation.src/services/alternatives/signals.tsandsrc/services/alternatives/factor-similarity.ts— candidate qualification and missing-source handling.