Momentum

A personal momentum scanner with AI-powered news synthesis. Built for one trader.

When a Fed pause hits the wire, a generic scanner shows you the headline. This shows you who benefits — with the mechanism, the confidence, and the horizon — alongside the price action that triggered the alert.

ticker scan → routing → news ingest → classify → synthesize → macro reasoning
                                                                      ↓
                                                           alerts + dashboard

Three signal classes drive alerts: threshold (watchlist + big moves), delta (rank changes between scans), and macro (events with second-order beneficiary breakdowns). All throttled, all auditable, all opinionated.

Mobile-first PWA with light/dark, sortable tables, per-ticker drill-down, and a separate macro-events view. Runs on its own — no babysitting.

---

What it does

Every scan, the pipeline:

1. Pulls daily bars + intraday snapshots for ~2,500 tickers (S&P 500 + NASDAQ-100 + NYSE/NASDAQ commons) from Alpaca
2. Computes technicals (RSI-14, MACD crossover, 1d/5d change, relative volume, sparkline)
3. Routes ~50–200 tickers to news ingest based on price/volume signals
4. Classifies news (Claude Haiku) → synthesizes per-ticker explanations (Claude Sonnet) → reasons about macro events (Claude Opus)
5. Builds tiered alerts (catalyst > macro > watchlist > delta) with per-ticker throttling
6. Delivers to Feishu webhook + commits scan/news/deltas JSON for the web dashboard

Architecture

Three-tier Claude pipeline with cost control:

Tier	Model	Job	Per-scan cap
0	Rules	Route ~2,500 → ~100 tickers (price/volume/delta filters)	—
1	Haiku 4.5	Classify news (type, impact, dedup, route flag) — batches of 15	100 items
2	Sonnet 4.6	Per-ticker synthesis (verdict, confidence, supporting news) — 8 concurrent	20 syntheses
3	Opus 4.7	Macro reasoning (beneficiaries, losers, mechanisms) — 4 concurrent	5 events

Tier-0 routing is the cost lever: rules filter 2,500 tickers down to ~100 before any LLM cost. Each tier uses tool-use for structured output, prompt caching to drop input cost ~90% on repeat system prompts, and full audit logs to data/audit/YYYY-MM-DD/.

Macro reasoning is the differentiator. Opus is constrained to the universe (drops unknown tickers) and required to give explicit causal links — "tariff on Chinese solar → First Solar margin expansion" — not vibes. Confidence levels (high/medium/low) and horizons (intraday/days/weeks/months) are part of the schema.

Delta tracking persists prior top-20 movers, detects rank jumps (≥15 positions), new entrants, and momentum acceleration — surfacing early movers before they become headlines.

Alert tiers:

• A0 catalyst: synthesis verdict says news explains the move + ticker also has a big move/unusual volume — always fires
• A threshold: watchlist (±1.5%+), big move (±3%+ RTH / ±5%+ AH with 2× volume) — capped at 5/scan
• B delta: new top-20, rank jumps, momentum acceleration — capped at 5/scan
• C macro: high/medium-impact macro events with beneficiary lists — always fires

Per-ticker/type throttling (2-hour cooldown) prevents alert spam.

Stack

Backend — Python 3.11+ · Alpaca paper API (bars + Benzinga news) · RSS (Fed, CNBC, MarketWatch, Yahoo) · Anthropic SDK · pandas/numpy · feedparser · httpx

Frontend — SvelteKit · Tailwind CSS · Vercel deployment · TypeScript

Ops — GitHub Actions cron-triggered scans · commits scan/news/deltas JSON to repo · Vercel auto-deploys frontend · Feishu webhooks for alerts (optional separate webhook for Chinese-language macro digest)

Web dashboard

• / — recommended picks (LLM-scored, with conviction), scan filters, top-20 movers, fresh-news tickers, watchlist, full sortable table — see Dashboard design below
• /t/[ticker] — per-ticker drill-down (full technicals, all news, synthesis)
• /macro — macro events with beneficiaries/losers
• /weekly — Saturday roll-up (top movers, catalysts, rank jumps by day)
• /political — disclosed Congress trades (Senate + House STOCK-Act PTRs from FMP; requires FMP_API_KEY). Cross-references the day's scan: tickers also in the watchlist or moving ≥ 3% get highlighted.
• /performance — past alert outcomes (1d/3d/5d returns)

Dashboard design

The scan page uses the Conviction visual direction — warm off-white canvas (#fafaf7) in light mode, dark slate in dark mode, with directional card tints (green wash for longs, red for shorts). Playfair Display is loaded for the hero % number and conviction score; everything else stays in the system sans / JetBrains Mono stack.

Sections, top to bottom:

• Recommended — hero PickCard grid (3 columns). Each card: rank · ticker (large bold) · long/short badge · price · Playfair display % · inset SVG sparkline · conviction score (1–10) · flag chips · synthesis "why" · top supporting headlines. Bucketed by long-term (catalyst-backed) vs short-term (pure technical), both sorted by score.
• Top 20 movers — dense TickerRow grid (rank · ticker · block-char sparkline · price · % · chips). Sorted by |%chg| of the filtered universe. Each row optionally carries a one-line "why" pulled from synthesis summary or top headline.
• Fresh news — tickers with new headlines that aren't already in Top 20.
• Watchlist — same thin-row treatment, ★ next to the ticker.
• All scan — collapsible <details> with the full sortable/searchable table.

Flag chips (replaces the old pill-badge stack — single normalized vocabulary in lib/flags.ts, rendered by IconCluster with up to 3 visible and a +N overflow chip):

Chip	Tone	Means
LATE	red	Late entry risk — price stretched well above its trend
NEWS	warn	High-impact news headline today
VOL	warn	Volume ≥ 2× 20-day average
NEW	info	New entrant to Top 20 since the last scan
JUMP	info	Big rank jump or accelerating momentum
MACD↑	up	MACD bullish cross
MACD↓	down	MACD bearish cross
OB	warn	Overbought (RSI > 70)
EXT	warn	Extended above moving averages
★	mute	On watchlist

Hover any chip for the full plain-English tooltip.

Sparkline strategy is hybrid — hero PickCards get a readable SVG polyline at full opacity (top-right inset of the % number); thin TickerRows get unicode block chars (▁▂▃▄▅▆▇) in JetBrains Mono. One Sparkline component handles both via a treatment: 'line' | 'block' prop.

Components (web/src/routes/): PickCard.svelte, TickerRow.svelte, IconCluster.svelte, Sparkline.svelte, FilterBar.svelte, ScanTable.svelte. Tokens live on :root / [data-theme='dark'] in app.css so the same components render correctly in both themes without per-component branching.

Running it

Required env vars in .env:

ANTHROPIC_API_KEY=...
ALPACA_API_KEY=...
ALPACA_API_SECRET=...
FEISHU_WEBHOOK_URL=...        # optional; alerts log only without it
FEISHU_MOM_WEBHOOK_URL=...    # optional; Chinese macro digest
FMP_API_KEY=...               # optional; powers /political dashboard
                              # (250 req/day free at financialmodelingprep.com)
SITE_URL=...                  # optional; base URL used in Feishu card links
                              # (defaults to the Vercel deployment)

# Backend scan
python -m scanner.main

# Web dev
cd web && npm run dev

# Web build (Vercel)
cd web && npm run build

Each layer fails independently: missing Anthropic key → LLM tiers skipped, scanner still returns technicals; missing Alpaca → no news/bars; missing Feishu → alerts log only.

---

Personal scanner. Not investment advice. Source by request.