Concepts
What is Trading Hub?
A research and analytics platform for publicly traded instruments — equities, ETFs, crypto, FX, rates, commodities, REITs. It is not a broker, not investment advice, and not a fiduciary. It produces probability-distributed forecasts + per-action playbooks + a public calibration scoreboard. See AI features.
Why is the named dimension "probability-first forecast + playbook + calibration"?
Bloomberg, TradingView, Koyfin, FactSet ship quotes + charts + economic data. None ship probability-distributed forecast cones with per-action playbooks AND publish their calibration. The named dimension is the entire stack.
Is this safe? Can I trust the outputs?
Every score is decomposable to the factors that drove it. Every external feed shows its source + freshness. Calibration is measured weekly via Brier score on /trust. Mistakes are documented within 7 days. That said: AI hallucination is possible, models drift, feeds break — verify critical claims independently. See Terms §6.
Scores & tiers
What's a tier?
A categorical buy/sell signal: Strong Buy / Buy / Hold / Sell / Strong Sell. Mapped from the 0-100 score by thresholds (e.g. score ≥ 80 → Strong Buy, ≥ 60 → Buy, ≥ 40 → Hold, ≥ 20 → Sell, < 20 → Strong Sell). Tiers exist because round numbers are easier to act on than raw composites.
"Tier" appears in different places — does it always mean the same thing?
No — there are three independent uses of the word "tier":
- Directional tier (Strong Buy → Strong Sell): the signal-strength tier above; appears on /security-deep, /expected-returns, /best-plays, /confluence, and the Today's Edge hero on /now.
- Urgency tier (critical / today / this-week / this-month): the time-sensitivity classification of an action item on the /now cockpit only. This says when something is actionable, not which way.
- Factor tier (1=macro, 2=sector, 3=instrument, 4=technical/sentiment): the layer a contributing factor lives in. Used in /factor-library and /factor-attribution.
What's a score?
A 0-100 composite from the multi-factor weighted scoring engine. Each factor contributes its weight × its z-score. See scoring model card. The score answers "how attractive is this security right now" on a normalized scale across 14 asset classes.
"Score" — what's the difference between composite score, mosaic strength, and prescience tier?
Three computationally distinct numbers:
- Composite score (0-100): weighted-factor attractiveness; the canonical "score." Drives the directional tier.
- Mosaic strength (0-100): multi-source agreement strength on /security-deep — built from up to 11 independent evidence sources (insider tape, 13F, options flow, news, technical, fundamentals, on-chain, macro, history, social, smart-money). Different from composite score because it measures cross-source corroboration, not factor-weighted attractiveness.
- Prescience tier: derived from the 5d-horizon GBM forecast after Platt + isotonic calibration + ensemble bandit weighting. Drives Today's Edge on /now. Different from composite tier because it measures probability of upward move in a specific horizon, not multi-factor attractiveness.
What's alpha?
Excess return per unit of time, in basis points (bps). Alpha > 8 bps qualifies for "buy" action; below 3 bps is below cost friction (slippage + spread eats the edge). Net alpha (`alpha.netBps`) accounts for execution drag.
What's the difference between score and alpha?
score is a categorical attractiveness measure (0-100). alpha is a continuous return-per-time measure (bps). Score is stable across regimes (a 70 means roughly the same thing in risk-on and risk-off). Alpha varies — bps in risk-off can mean very different things from bps in risk-on.Which DECIDE page do I use?
Per usability audit Win 5 (2026-05-05). The DECIDE section has 6 pages each ranking tickers by a different metric. They are ORTHOGONAL — a ticker can be #1 on one page and #15 on another. Use the right page for the question you're asking.
| Page | Question it answers | Key metric | When to use |
|---|---|---|---|
| /now | What should I do RIGHT NOW? | composite score + tier | Daily action surface; pre-market routine |
| /expected-returns | Which tickers have best log-wealth growth? | g = E[r] − ½σ² | Long-term wealth optimization (Itô-corrected) |
| /best-plays | Which options plays have highest information edge? | Kelly-weighted edge (bps) | Tactical options trade selection |
| /confluence | Are signal frames AGREEING or CONTRADICTING? | independence_score + state | Conflict resolution (Now-buy vs Security-deep-sell) |
| /diversifiers | Which tickers improve my portfolio's risk-adj return? | DAR = μ / (σ × ρ̄) | Portfolio fit (NOT directional alpha) |
| /hold-through | Which positions to hold through THIS session? | g_window (session-relative) | Intraday holding decisions |
| /factor-attribution | Why did this score get this tier? | factor weight × z-score | Score forensics + trust calibration |
Power-user workflow: /now (action) → /security-deep (drill) → /factor-attribution (why) → /confluence (cross-frame check) → /best-plays (options trade) → /diversifiers (portfolio fit) → /hold-through (intraday). Each is orthogonal — a ticker ranking #1 on /best-plays might rank #15 on /diversifiers. That's NOT a bug; it's how independent metrics should behave.
DISCIPLINE side: after acting, check /cohort-brier + /calibration-scoreboard + /trustee-report for engine reliability self-audit.
Factors
What's a factor?
A single signal contributing to the composite score. THv2 has ~30 factors per asset class. Each is tagged with a tier (1=macro, 2=sector, 3=instrument, 4=technical/sentiment) and a weight. Examples:
F05_mom3m (3-month momentum), F12_breadth (sector breadth), F23_liquidity (spread + book depth), F26_insider (insider buying).What do the factor IDs mean?
Format:
F{NN}_{label}. Tier prefix is the leading digit:
- Tier 1 F0X — macro / cross-asset (e.g. F01_macro_mood)
- Tier 2 F1X — sector / industry (e.g. F12_breadth)
- Tier 3 F2X — instrument fundamentals (e.g. F22_earnings, F23_liquidity)
- Tier 4 F3X+ — sentiment/technical (e.g. F31_funding, F32_news_velocity)
Why don't I see all 30 factors on every score?
UI shows the top 5 contributors by absolute value (Bible §6.5 progressive disclosure). The full decomposition is available in the score record on
/security?ticker=X. Hidden factors aren't ignored — they affect the composite, just don't headline.Are factor weights fixed?
Layer-conditional: weights vary by regime per LS6 (engine v76). Risk-on rewards momentum (F05_mom3m); risk-off rewards quality (F22_earnings, F26_insider). Same security can have different scores in different regimes.
Regime
What's a regime?
A label for the current market state. THv2 uses three primary states: risk-on (broad participation, momentum dominant), risk-off (defensive, quality dominant), transition (whipsaw zone, low confidence). Computed from cross-asset signals (DXY, yield curve, breadth, vol regime, credit spreads).
How long does a regime last?
Half-life by state: risk-on 42 days, risk-off 28 days, transition 10 days. After the half-life, persistence probability drops below 50%. Visible on /security as "P-persist-5d" — probability the regime lasts another 5 days given it's been in this state for N days.
Why does a "transition" regime matter?
Whipsaw zone — first 24-48h of a regime flip can show contradictory signals. The playbook engine refuses to produce action signals during transition (per "when NOT to act" disqualifier). Wait for the regime to settle.
Forecast cones
What's a p10/p50/p90 cone?
A probability-distributed forecast. p50 = median expected outcome (50% of paths land here or worse). p10 = 10th percentile (only 10% of paths are worse than this — the lower bound). p90 = 90th percentile (only 10% beat this — the upper bound). The cone widens with horizon and volatility. See prescience model card.
What's a half-life?
How long it takes for the predictable component of an expected move to decay by half. Short half-life (1-2 days) → fast mean-reversion; the signal is gone quickly. Long half-life (10+ days) → slow trend; the signal persists. The playbook engine rejects very short half-lives because retail-speed execution can't catch them.
Why doesn't the forecast match what actually happened?
It often won't — that's why we publish a calibration scoreboard. If realized lands within p10/p90, the cone was honest. If it lands outside, the variance estimate was wrong. We track this weekly via Brier score; engines below 0.55 reliability get re-weighted. See /trust.
Playbook
What's R:R?
Risk-to-reward ratio.
R:R = |target - price| / |price - stop|. R:R ≥ 2.0 is a strong setup (you're betting $1 to make $2+). Below 1.5, the playbook downgrades to "wait" — even if the signal is right, the bet doesn't pay enough.Why does the playbook say "wait" when the score is high?
Six "when NOT to act" disqualifiers can override even a Strong Buy: low confidence (<40), tiny alpha (<3 bps), thin liquidity (F23 <25), short half-life (<5 bars), extended hours, or transition regime. See playbook model card.
What's quarter-Kelly sizing?
Full Kelly = optimal bet size if your edge is exactly known. Quarter-Kelly = 25% of full Kelly. Trades ~75% of expected long-run growth for ~6% of full Kelly's variance. Also capped at 15% per position.
Alerts
What triggers an alert?
Four kinds: score-jump / score-drop (≥ 8 pt move within 30s), tier-transition (Hold→Buy etc), lead-indicator firing (one of 6 detectors trips), regime-flip (risk-on ↔ risk-off ↔ transition). Severity: info / warn / strong drives badge color.
Why am I not seeing alerts?
Pipeline runs every 30s in the foreground tab. Background tab → throttled. If you've been on the home page for > 5 min and there's no inbox activity, the universe might just be quiet — check /status to verify the pipeline is firing.
Data & freshness
What does the freshness badge mean?
Five states: fresh (within TTL — green), stale (1-3× TTL — amber), degraded (>3× TTL — red), conflict (two feeds disagree — purple), unknown (no timestamp — gray). Hover for source, exact ts, ttl.
Where does the data come from?
Twelve public feeds: CoinGecko, Yahoo Finance, FRED, SEC EDGAR, FDA, EIA, IMF, DeFi-Llama, Tradier, Binance Futures, COT, Coinbase Exchange. Full attribution on /trust. The moat is how 12 feeds are recombined, not access to private data.
How fresh is "fresh"?
Per-source TTL: spot crypto 5min, equity 60s, macro 1h, filings 1h, FX reserves 24h, COT 7d. Listed in scoring model card.
Privacy
Do you store my data?
No — there are no user accounts. Your watchlist, journal, settings live in your browser's localStorage and IndexedDB. We see only Vercel platform request logs (30 days, no PII tied) and AI prompt logs (when you use AI features, 30 days). See Privacy Policy.
How do I export / delete my data?
Export: command palette (Cmd+K) → "Export local data (CSV)" — gives JSON of watchlist + journal + settings. Delete: command palette → "Clear all caches + reload" — wipes everything client-side. Server-side AI prompt logs: email kadeem.yearwood@gmail.com for deletion (30-day SLA).
Troubleshooting
The app feels frozen / scores aren't updating
Check /status first. If services are green but scores are stale: command palette → "Clear all caches + reload" forces fresh state. If feeds are red, that's external — wait or use the cached scores.
I see different prices on Trading Hub vs my broker
Yahoo + CoinGecko data has 1-15 min delay depending on instrument. Broker is real-time. Trading Hub doesn't optimize for real-time — it optimizes for forecast + playbook on 1-30 day horizons. Use broker for execution timing.
An alert fired but the move didn't materialize
Alerts fire on detected moves AT the moment of detection. Mean-reversion can pull the move back before you see it. Per playbook, half-life < 5 bars disqualifies action — alerts on those are informational only.
I want to report a bug or contribute a research idea
Email kadeem.yearwood@gmail.com. For security issues: /.well-known/security.txt. Bug reports get acknowledged within 48 hours.