QuantPulse · Live US Equity Quant Terminal

A live-updating, zero-build quantitative research terminal covering 38 US mega-cap equities across all 11 GICS sectors with 11.7+ years of real daily market data (Jan 2015 → snapshot, split-adjusted, via Yahoo Finance) — plus a Netlify serverless function that streams fresh quotes so the dashboard never goes stale.

Everything — live merging, charting, factor analytics, machine learning, backtesting and risk statistics — runs in your browser. No build step, no API keys, no tracking, no ML libraries.

What's new vs. a static bundle

Capability	How it works
Real-time data	netlify/functions/quotes.js fetches Yahoo Finance server-side (no CORS) and returns compact JSON. The client polls every 15–60s, merges completed sessions into the analytics history, and carries today's session as a forming bar (live price line, never contaminating backtests).
Graceful degradation	If the function is unreachable (e.g. opened as a plain static site), the app badges itself SNAPSHOT and keeps every feature on the bundled data.
Pro charting	Candlestick / line / area charts with wheel zoom, drag pan, pinch zoom, full crosshair with axis tags, OHLC legend readout, SMA 20/50/200 + EMA 50 + Bollinger overlays, Volume / RSI-14 / MACD panels, log scale, 1D-5m and 5D-15m intraday, peer comparison and relative strength.
Live screener	Sortable 38-name cross-section: price, day/1W/1M/3M/YTD/1Y returns, 20D vol, RSI-14, % off 52W high, volume ratio 20/120, 30-day sparklines. Search + sector filter, click-through to the Terminal.
Factor lab	7 classic factors + a composite (average z-score): live cross-sections, client-recomputed rank-IC history with t-stats, quintile spreads (Q1−Q5), 36-month factor correlation matrix. IC history auto-extends as live sessions merge.
Backtester	Position-based engine with daily weight drift, long-only & long-short books, turnover-based costs, 12 performance metrics (Sharpe, Sortino, Calmar, alpha/beta, skew, turnover…), monthly heatmap, live holdings, CSV export.
ML lab	Four models written from scratch on Float32Array maths — ridge regression (closed form), elastic net (coordinate descent down a lambda path), a one-hidden-layer neural network (ReLU + Adam, mini-batch) and gradient boosting (histogram regression trees) — plus their equal-weight ensemble. Nine z-scored features predict next-month returns through a purged walk-forward protocol (training rows whose 21-day label overlaps the test date are embargoed), evaluated with out-of-sample rank ICs, quintile spreads, hit rates, permutation feature importance, a model scoreboard against the no-ML composite baseline, and strategy backtests on the same position engine as the Backtest tab. Seeded RNG → every run is reproducible; a full 4-model benchmark trains in ~12 s in the browser.

Deploy to Netlify (drag & drop)

The bundle is intentionally deployable as-is:

1. Zip this folder (or push it to GitHub and connect it).
2. Drag & drop at app.netlify.com/drop — netlify.toml pre-configures everything: static publish + the quotes function + the /api/quotes redirect.
3. Done. The header badge flips to LIVE within seconds.

Netlify Functions (free tier includes 125k invocations/month) powers the live feed. The 20-second server-side cache keeps upstream request volume low when several clients poll at once.

Any other static host works too — you just lose live updates and the app badges itself SNAPSHOT (everything else still runs).

Run locally

cd quantpulse
python3 -m http.server 8080        # static only → SNAPSHOT mode
# or, with the live API mirrored locally:
node dev-server.js                 # optional helper (serves /api/quotes)

Project structure

quantpulse/
├── index.html                    # single-page terminal shell
├── netlify.toml                  # publish + functions + /api redirect
├── netlify/functions/quotes.js   # live data proxy (Yahoo → JSON, cached)
├── css/style.css                 # dark terminal theme
└── js/
    ├── data.js                   # bundled market snapshot (~1.7 MB)
    ├── divs.js                   # bundled ex-dividend calendar since 2015
    ├── charts.js                 # zero-dependency canvas chart engine
    ├── live.js                   # polling, market clock, pub/sub, fallbacks
    ├── core.js                   # live-merge engine, factors, IC, backtest,
    │                             #   risk + price/total return basis switch
    ├── ml.js                     # from-scratch ML: ridge / elastic net / MLP /
    │                             #   gradient boosting, purged walk-forward
    └── app.js                    # 7 modules: Overview / Terminal / Screener /
                                  #   Factors / Backtest / ML Lab / Risk

Data & methodology notes

• Prices are split-adjusted daily OHLCV from Yahoo Finance; volumes in the bundle are stored in thousands.
• Live merge rule: a bar dated today counts as completed only after 16:05 ET; earlier it is a forming overlay excluded from all analytics.
• Ex-dividend-aware split detection: the live feed carries Yahoo's split & dividend event calendar. When a split's ex-date falls after the snapshot, the ticker's entire bundled history is rescaled to the new share basis (prices × den/num, volumes × num/den) before returns are chained — a 10:1 split never shows up as a fake −90% day, and split events that arrive after their bar has merged are repaired retroactively (including the equal-weight benchmark re-chain). Dividend ex-dates stay in the price chain (price-return convention) and are used to avoid mistaking an ex-div gap for a split; unexplained boundary gaps are re-checked against a 1-year event calendar, and when evidence is inconclusive the app warns instead of guessing. Only if the feed is down entirely can a strict open+close+volume heuristic apply an adjustment — always flagged [unverified].
• Factors are computed client-side from the merged history (momentum 12-1/6M, 1M reversal, 20D low vol, RSI-14 Cutler variant negated, 52W-high proximity, 20/120D volume trend). Rank IC = Spearman ρ between month-end scores and the next 21 trading days' returns; the Factors tab shows live-recomputed ICs (falling back to the offline bundle where candle volume history is unavailable).
• Backtests charge turnover × cost_bps per rebalance, fill at the rebalance day's close, and model daily weight drift within holding periods. No borrow/financing or slippage.
• ML methodology: features are winsorised at ±3σ and z-scored within each month's cross-section; targets are next-21-trading-day returns; every month-end the model retrains on the trailing window (36/60/96 months) and scores the current cross-section strictly out-of-sample — training rows whose label window has not fully realised before the test date are embargoed (purged walk-forward, no leakage). Out-of-sample results on a 38-name mega-cap panel are honestly thin (OOS ICs of a few hundredths with t-stats under 1): the lab is a methodology showcase, not an alpha claim.
• Return basis switch (PRICE / TOTAL): the header toggle selects the return convention for analytics. PRICE chains raw closes (cash dividends appear as ex-date price drops). TOTAL adds each cash dividend back on its ex-date — (close + div) / prevClose − 1, the standard approximation to a dividend-reinvested index — which is the honest view for high-yield names (e.g. PFE's 2015→2026 CAGR moves from ~7% to ~13% when its ~4-6% yield is counted). The basis propagates to risk statistics (vol, Sharpe, Sortino, VaR/CVaR, beta, correlations, drawdowns), backtests and their benchmark, IC / quintile forward returns, ML training targets, screener windows and the equal-weight universe index. Candlesticks, Terminal price charts and factor signals stay price-based (a candle cannot embed a cash payout; momentum-type signals are classically price-based; the low-volatility signal chains raw closes so no ranking shifts with the basis). The dividend calendar (1,400+ ex-dates since 2015, bundled in js/divs.js) extends live as new ex-dates merge.
• Dividend-uplift chips, sortable tables & yield filter: the risk table's CAGR and Sharpe cells carry an amber chip with the full-sample dividend uplift (TOTAL − PRICE, e.g. PFE +6.3pp CAGR / +0.24 Sharpe), so toggling the basis makes the change visible without memorising both views — sort by Div yld (PFE / AMT / DUK surface on top) and switch TOTAL to compare high-yield names at a glance. The screener gained a trailing-12-month Div yld column (sortable, ≥3% highlighted) plus a minimum-yield filter (Any / ≥1% … ≥5%; ≥3% keeps exactly the 7 payers PFE, AMT, DUK, CVX, PLD, HD, NEE), and both tables show a sort-direction arrow plus aria-sort on the active column header.
• Backtest dividend annotation: the Backtest lab's CAGR and Sharpe cards carry the same chip, computed by rerunning the engine on the other basis's return chain with the identical holdings path, weights and costs — so the chip is the strategy's dividend contribution (e.g. momentum long-only +2.1pp CAGR; equal-weight universe +2.7pp). Long-short books can show a red negative chip because the short legs owe the ex-dividend drop under the total-return convention.
• The universe is a static demonstration panel, not a tradable index.

License

MIT — see LICENSE.