Swing Desk

Single-ticker research dashboard for swing trading (holds of days to a few weeks). BUILD_PROMPT.md is the spec, and MANIFESTO.md overrides it on philosophy and design.

Status

All seven phases are built:

• 1 Engine: FastAPI, yfinance + FRED clients, SQLite cache, backoff, coverage script
• 2 Shell: layout, ticker search, watchlist, candles + volume + MA20/50/200 + RSI/MACD panes, TradingView embed
• 3 Earnings (surprise vs reaction, hold-window warning), fundamentals (divergence flags), valuation (own-history percentiles)
• 4 Options (IV term structure, skew, expected move), macro (FRED), regime badge in the top bar
• 5 Sentiment (VIX percentile, crowd signals), trap detector (+ headline tagging), verdict scorecard → "Log as trade"
• 6 Trade journal + behavior analytics (setup, emotion, weekday, plan-following, disposition effect)
• 7 Backtester (no look-ahead, tested), S1-S4, IS/OOS hold-out, walk-forward, variant counter, t ≥ 3 hurdle, CSV upload

Not yet run on live data. The build sandbox cannot reach Yahoo, so everything was tested against synthetic data (49 engine tests). Your first local run is the real test; tell me what breaks.

Pages

Page	What it is for
/t/TICKER	All panels for one stock, ending in the Verdict: score, thesis, invalidation, exit, "Log as trade"
/journal	Every trade, plus where your results come from (setup, emotion, weekday, plan-following, hold times)
/lab	Backtest S1-S3 (+ S4 regime filter) on tickers or an uploaded CSV
/settings	Scorecard weights (saved in your browser)

How it runs (no computer needed)

GitHub Action (weekdays after the US close)          Vercel website
  watchlist.txt -> engine/scripts/snapshot.py -> data branch -> reads JSON
GitHub Action "Run backtest" (you click Run) -> data branch -> Lab "Load latest result"

• Data: the Nightly data snapshot Action runs every weekday at 22:37 UTC. It runs every analysis for the tickers in watchlist.txt and publishes JSON to the data branch. To add a ticker, edit watchlist.txt on GitHub (max 60); a fresh snapshot starts automatically when you commit the change (~5-10 min), or run the Action by hand: Actions -> Nightly data snapshot -> Run workflow.
• Macro + regime badge: add a free FRED key as a repo secret named FRED_API_KEY (Settings -> Secrets and variables -> Actions). Without it those panels stay empty.
• Journal: stored in your browser. Use Export on the Journal page to back it up.
• Backtests: on the Lab page click Prepare GitHub run, copy the request, paste it into Actions -> Run backtest -> Run workflow, then Load latest result a few minutes later.
• Vercel: Root Directory web. Leave NEXT_PUBLIC_ENGINE_URL unset (setting it switches the site to live-engine mode).

The site reads https://raw.githubusercontent.com/expowolf/startupideas/data/..., so the repo must stay public. A private repo would need the data published elsewhere.

Optional: live engine on your computer

Any ticker on demand, journal in SQLite, CSV uploads in the lab. Windows commands:

cd engine
py -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
copy ..\.env.example .env          # add FRED_API_KEY
uvicorn app.main:app --port 8000

cd ..\web
npm install
echo NEXT_PUBLIC_ENGINE_URL=http://localhost:8000 > .env.local
npm run dev                         # http://localhost:3000

Engine tools (from engine/): python -m scripts.validate (data coverage report), pytest (offline tests).

Decisions log

Topic	Decision	Why
Hosting	Vercel UI + nightly GitHub Action snapshot (live engine optional)	Swing trading needs daily bars, not live quotes; free, nothing left running. Yahoo blocks many cloud IPs, so the Action fails loudly if it gets no prices
"Hedge fund algorithms"	Not built as such	Proprietary. We implement the published anomalies they trade (S1-S3)
Manipulation detection	"Trap detector" from daily OHLCV + yfinance headlines (phase 5)	Spoofing needs paid order-book data. Panel measures failed breakouts, stop sweeps and faded news gaps; it does not claim who caused them
TradingView	Free embed widget as a second chart tab	No account API exists. Widget data is display-only and never feeds the scorecard or backtests
FinanceDatabase	Not needed after all	Sector → ETF mapping is a fixed 11-row table (Yahoo sector → SPDR fund); a 100 MB dependency was not worth it
Regime rule	SPY vs 200-day average + VIX (25 / 20 thresholds), fixed in advance	High-yield spread shown but excluded: FRED only serves ~3 years of it, so it can't be backtested
Revenue-surprise filter (S1)	Not offered	yfinance has no history of revenue estimates
Yield curve	Added DGS1, DGS5, DGS30 to the FRED list	Three points don't make a curve; scripts.validate checks them like the rest
investing-algorithm-framework	Not adopted	Lacks the variant counter, IS/OOS hold-out and t-stat hurdle this spec requires
Congress trades	Not built yet (pending your call)	See research below
shadcn/ui	Components hand-written in shadcn style	The shadcn registry was unreachable from the build sandbox; run npx shadcn add … locally if you want the originals

Congress trading research (kadoa-org/congress-trading-monitor)

• Data is a static JSON snapshot (~157 MB) committed to their repo, built from House Clerk, Senate eFD and OGE filings. The repo has no scraper; Kadoa collects the data with its paid service, and "continuous updates" require contacting Kadoa sales.
• In its latest 5,000 trades, the median delay from trade to disclosure is 61 days, and 59% were flagged late.
• For a 2-week swing hold, that is history, not a signal.