QuantPulse

A quant factor research tool. Pick a stock, pick a factor, and find out whether it actually predicts the next move — measured the way a quant fund measures it, using real market data.

An introduction to quant research, built as a runnable Next.js app.

Running it

npm run dev

Then open http://localhost:3000.

Node runtime. This machine had no Node installation, so a portable build (Node 24.21.0) was extracted to %LOCALAPPDATA%\nodejs-portable and added to your user PATH. A new terminal picks it up automatically. If a shell still cannot find node, run dev.cmd in this folder — it pins the path first.

Command	What it does
npm run dev	Dev server on port 3000
npm run build	Production build
npm run start	Serve the production build
npm run test	Bar hygiene, pre-close window, sync merge, edge model, trade record and options-maths suites
npm run check	typecheck + lint + test + build
npm run deploy	Pre-deploy gate, then commit and push (Git Bash)

Pages

Route	Purpose
/	Landing page
/dashboard	The tool. Opens in Simple mode; Advanced holds the statistics
/screener	Every strategy against every ticker, filtered by strategy, company and risk
/watchlist	Saved ticker + strategy pairs, scanned in one pass
/notifications	Entry and exit changes across the watchlist
/settings	Appearance — accent, direction colours, background, grid
/guide	How to use it, and how to read every statistic
/membership	Account area and tiers (UI only — see below)
/api/history	Daily OHLCV for one ticker (GET)

Market data is real

Prices come from Yahoo Finance via yahoo-finance2 — the same endpoints the Python yfinance package reads, in TypeScript so the app stays on one runtime and deploys to Vercel without a Python function.

Any ticker Yahoo carries works: AAPL, BRK-B, BTC-USD, ^GSPC, VOD.L, EURUSD=X. Three years of daily bars, fetched once per symbol and cached at the edge for an hour.

src/lib/market.ts is server-only (yahoo-finance2 needs node:module and cannot be bundled for the browser). Anything the client needs — the ticker list, shared types — lives in src/lib/symbols.ts.

The landing page reads live

The hero panel on / is not a screenshot. It runs the real engine on AAPL at request time and prints the current signal, dated. The page is prerendered with revalidate = 3600, so it refreshes hourly and costs one Yahoo call an hour no matter how much traffic it gets. If the provider is unreachable at build or revalidate time it falls back to describing the tool rather than showing numbers that were never true.

Live context sits beside the algorithm, never inside it

The terminal also shows the extended-hours price and recent headlines, because a signal computed on the daily close cannot see an earnings release at 16:05. When the after-hours move exceeds 2%, the signal is flagged as possibly out of date.

None of it touches the maths, for two reasons:

• No history to test against. Yahoo returns roughly half a day of headlines. There is nothing to fit or validate a news factor on across three years of prices.
• It would be lookahead bias. Mixing today.s news into a historical simulation is the exact error the rest of the tool is built to detect.

Headlines are labelled as context, not cause. A story published near a price move is a hypothesis; the interface says so rather than asserting a reason. Building a real news factor would need point-in-time historical news with timestamps and sentiment — a different data provider, not Yahoo.

Screener

/screener answers the question the terminal cannot: which ticker should I even be looking at? It runs all five price factors against the universe at their default settings and lets you filter the results three ways:

Group	Filters
The strategy	verdict, factor, signal now, minimum trades, beat holding out of sample, survives the stress test, hide strategies that still lost money
The company	sector, industry, market cap, P/E, dividend, profitable or not, revenue growing or shrinking
Risk and tradeability	liquidity (daily dollar volume), volatility, beta, short interest

Simple mode shows four of these; the rest live behind More filters in advanced mode. Every bucket carries its ticker count (Mid · $2–10B (46)) so an empty slice announces itself rather than silently returning nothing.

Liquidity is the one that matters most and the one a conventional screener never has. Every backtest here charges a flat 10 bps round trip, which is only honest on a name that actually trades; below roughly $10M a day the real spread is wider than the modelled cost and the edge shown is partly fiction.

Analyst ratings and price targets are deliberately absent. They are opinions, not measurements, and nothing else on this page is an opinion.

It is precomputed, not live. The scan is a thousand backtests and takes about three minutes — far too slow per request, and pointless, because daily bars only move once a day. scripts/build-screen.mts writes public/screen.json (399 KB, ~62 KB gzipped) and the page fetches that static file. No Yahoo traffic, no function timeout, nothing added to the JS bundle.

Rebuild it after the close:

node scripts/build-screen.mts

Two rules the script enforces, both of which a conventional screener breaks:

This is the beginning of the README. Read the rest at the original. Read the full README ↗