Free Quant Platform

An autonomous stock-analysis and paper-trading platform built entirely on free, public data: no Bloomberg, no paid APIs, no API keys.

Source	Used for	Access
Yahoo Finance (yfinance)	Daily + 15-min OHLCV, VIX, indices, quotes	free, ~15 min delayed
FRED, St. Louis Fed	Inflation, Treasury yields, employment, credit spreads	pandas-datareader, falling back to FRED's public CSV
SEC EDGAR (data.sec.gov)	10-K / 10-Q financials via XBRL	official free API
Google News, Yahoo Finance, CNBC, MarketWatch RSS	Headlines	feedparser + BeautifulSoup
FinBERT (ProsusAI/finbert, HuggingFace)	Financial sentiment	open-source model, runs locally
Kenneth R. French Data Library	Fama-French 3 factors + momentum, daily	pandas-datareader, falling back to the library's zipped CSVs

Paper trading only. Nothing here places real orders, and nothing here is investment advice.

Project layout

quant-platform/
├── requirements.txt
├── requirements-quantum.txt   optional: pinned Qiskit for the website's quantum simulations
├── .env.example
├── quant/
│   ├── config.py              settings, watchlist, FRED series                 [P1]
│   ├── db.py                  engine, sessions, SQLite/Postgres upsert          [P1]
│   ├── models.py              ORM tables                                        [P1, extended in P4]
│   ├── ingestion/
│   │   ├── http.py            retrying HTTP sessions                            [P1]
│   │   ├── market_data.py     yfinance OHLCV + quotes                           [P1]
│   │   ├── macro_data.py      FRED                                              [P1]
│   │   ├── sec_edgar.py       EDGAR 10-K / 10-Q metrics                         [P1]
│   │   ├── news_scraper.py    RSS scraping, sentiment pass, news-to-chart map   [P1]
│   │   ├── sentiment.py       FinBERT with a lexicon fallback                   [P1]
│   │   ├── relevance.py       company-name relevance filter for headlines       [P2]
│   │   ├── factors.py         Fama-French factor returns                        [P3]
│   │   ├── index_membership.py point-in-time S&P 500 membership (Wikipedia)     [P6]
│   │   └── pipeline.py        orchestrator + CLI                                [P1]
│   ├── analytics.py           50/200 MA, RSI, MACD, Bollinger, regime, anomalies [P2]
│   ├── quant_models.py        Fama-French 3F, momentum, mean reversion, scoring [P3]
│   ├── trading_bot.py         signals, paper ledger, Kelly/vol sizing, stops    [P4]
│   ├── scheduler.py           APScheduler daily run                             [P4]
│   └── web/                   static website: export, publish, quantum simulations
├── dashboard/                 Streamlit + Plotly                                [P5]
│   ├── app.py                 entry point and navigation
│   ├── common.py              cached data access, palette, chart styling
│   └── pages/                 market_overview · stock_deep_dive · financials_research · trading_bot
├── data/                      quant.db (SQLite) and cache/
└── tests/

The dashboard is Streamlit + Plotly, so the whole platform is one Python codebase reading one database.

Setup

python -m venv .venv
.venv\Scripts\activate            # macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cpu
copy .env.example .env            # then set SEC_USER_AGENT to "Your Name [email protected]"

Phase 1: running the ingestion engine

python -m quant.ingestion.pipeline                         # everything
python -m quant.ingestion.pipeline --steps prices macro    # a subset
python -m quant.ingestion.pipeline --tickers AAPL NVDA --no-finbert
python -m quant.ingestion.pipeline --steps fundamentals --refresh-fundamentals  # re-extract after adding a metric
python -m pytest                                           # offline tests

The first run backfills 10 years of daily bars and downloads FinBERT (~440 MB). After that, each run fetches only what's new: recent bars, revised FRED prints, filings it hasn't seen, and fresh headlines.

What each step stores

Table	Contents
price_bars, stock_splits	1d bars keyed by session date; 15m bars keyed by UTC bar start; split history
factor_returns	Daily Fama-French Mkt-RF, SMB, HML, RF and momentum (decimals), lagging today by 1-2 months
macro_series, macro_observations	FRED metadata and values (daily, weekly or monthly, depending on the series)
companies, fundamentals	One row per 10-K / 10-Q (about 11 years): revenue, net income, EPS, diluted shares, assets, liabilities, equity, total debt, cash, operating cash flow, capex, FCF, D/E, net margin, link to the filing
news_articles	Headline, publisher, UTC time, FinBERT label/score, and its intraday chart alignment
index_members	One row per stint of a stock in the S&P 500: ticker, name, added, removed (only with --universe sp500; see the research-universe section)

Design notes

• SEC metrics come from XBRL, not HTML scraping. Every fact in EDGAR's companyfacts API carries the accession number of the filing that reported it, so metrics are matched to an exact 10-K / 10-Q. A 10-Q reports both quarterly and year-to-date figures; the parser takes the quarter for the income statement. Cash-flow statements in a 10-Q are year-to-date only, so cash_flow_period_days records the length. Total debt is approximated from the standard debt tags, with guards against double counting.
• News-to-chart mapping. Each headline is aligned to the first 15-minute bar that traded after publication. Overnight news aligns to the next open, so the gap is captured. reaction_return is the move from the last close before the headline to the end of a 60-minute trading window. Market-wide headlines are mapped onto SPY.
• Sentiment score = P(positive) − P(negative) from FinBERT, in [−1, 1]. If torch / transformers are missing, a finance lexicon scorer takes over, and the sentiment_model column records which scorer produced each score.
• Resilience. Each ticker and series runs in its own transaction, every write is an idempotent upsert, and HTTP calls retry with backoff. SEC requests are throttled below its 10 requests/second limit.
• User agents matter. Yahoo's RSS returns 429 to non-browser agents, while FRED stalls agents that claim to be browsers. http.py sets each one accordingly.

Phase 2: analytics

python -m quant.analytics                       # current market regime
python -m quant.analytics AAPL JNJ              # + indicators and explained daily anomalies (90 days)
python -m quant.analytics TSLA --interval 15m   # intraday anomalies (30 days)

Everything is computed on demand from the database. The core functions in quant/analytics.py take plain DataFrames, so later phases can reuse them for backtests.

• Indicators (add_indicators, via the ta library): 50/200-day SMA with golden/death-cross events, RSI(14), MACD(12,26,9), Bollinger Bands(20,2), ATR(14) and 20-day realised volatility.
• Market regime (market_regime_for), computed for every day of history:
  • Trend uses the S&P 500. Bull when the 50-day MA is more than 1.0% above the 200-day and price is above the 200-day. Bear for the mirror image, or when price is under the 200-day and at least 20% below its 52-week high (this catches crashes before the lagging death cross). Otherwise Sideways.
  • Volatility uses the VIX: Calm below 16, Volatile at 24 or higher.
  • The resulting labels are Bull/Bear Trending, Bull/Bear Volatile, Sideways and Sideways Volatile.
  • Fast to de-risk, slow to re-risk: a label moves to a riskier state at once, but back to a calmer one only after 5 consecutive calmer days. This applies to trend, volatility and stance alike.
  • Macro overlay from FRED: inverted 10Y-2Y yield curve, the Sahm-rule recession signal, credit stress, and core CPI above 3%. Credit stress means the Baa-Treasury spread is at least 3.0% or has widened at least 0.5pp in 3 months. These feed a 0–1 risk score that sets the stance (Risk-On below 0.33, Risk-Off above 0.6). Each FRED series is shifted by its publication lag (45 days monthly, 7 days weekly), so the history never uses data before it was public.
• Chart-to-news correlation (explain_anomalies):
  • Flagging: a daily move of at least 8%, or one of at least 3σ and at least 2%, where σ is a robust volatility estimate (scaled median absolute deviation) so one crash can't hide the aftershocks. That's about 9 markers per stock per year. Intraday 15-minute bars use 3% / 4σ / 1.25%, and the opening bar, which carries the overnight gap, is scored only against previous opens.
  • Move window: from the previous close to this close, so overnight earnings count.
  • Driver: each move is classed as market-driven or stock-specific, using beta against SPY.
  • Ranking: candidate headlines are ranked by relevance × timing (before, during or after the move) × FinBERT agreement with the direction of the move.
• Relevance filter (ingestion/relevance.py): scores each headline 0–1 for whether the company is its subject. It uses curated aliases plus the SEC registrant name, and catches name variants ("Johnson and Johnson").
  • Demoted (0.4): multi-company lists, week-ahead previews, market roundups, and headlines led by another company's stock.

Calibration

All Phase 2 thresholds are reproducible with scripts/calibrate_phase2.py. The goal is robust settings (plateaus, stability, clear separation of forward risk), not maximum backtest returns, so Phase 4 backtests on the same history aren't silently overfit.

python scripts/calibrate_phase2.py regime          # trend band, drawdown rule, VIX cut-offs (1990-today)
python scripts/calibrate_phase2.py regime-confirm  # confirmation window: switches/yr vs detection lag
python scripts/calibrate_phase2.py regime-stance   # risk-score stance cut-offs, credit-stress input
python scripts/calibrate_phase2.py anomalies       # move rules vs SEC 8-K Item 2.02 earnings dates
python scripts/calibrate_phase2.py relevance       # headline filter vs the hand-labeled fixture

Setting	Before	Tuned	Evidence (1990–2026 unless noted)
Trend band / confirmation	1.5% / none	1.0% / 5 days	trend switches 6.2 → 4.2/yr, median spell 8 → 22 days, bear detected 6 days sooner
VIX spike rule	on	removed	forward vol on spike days 16.7% vs 16.2% at the same VIX level (no signal)
Volatility confirmation	none	5 days	switches 24 → 11/yr; Volatile still carries 2.2× Calm forward vol
Credit-stress input	HY OAS (3 yrs on FRED)	Baa−10Y spread	coverage 12% → 100% of days; stressed days' forward vol 20.7% vs 11.9%
Stance confirmation	none	5 days	stance switches 7.0 → 4.2/yr; Risk-Off forward vol 23.3% vs 10.9% Risk-On
Daily anomaly rule	5% / 3σ / 2%	8% / 3σ / 2%	per stock 13.8 → 8.7 flags/yr; busiest/calmest 5.8× → 2.9× (2016–2026)
Intraday opening bar	shared σ	own σ	opening-gap flags 0.30 → 0.04 per ticker-day
Intraday minimum move	1.0%	1.25%	TSLA 16 → 3 flags in 30 sessions, others barely change (provisional: small sample)
Relevance filter		7 rule fixes	precision 85.4% → 99.0% (568 labels)

Phase 3: quant models

python -m quant.ingestion.pipeline --steps factors   # daily Fama-French factors
python -m quant.quant_models                         # exposures, backtests, signals, composite scores
python -m quant.quant_models --model carhart         # add the momentum factor to the regressions
python -m quant.quant_models --evaluate              # monthly score history + information coefficients

quant/quant_models.py loads prices, filings, splits and factors once into a Universe, so scores can be recomputed cheaply for any past date. That is what Phase 4 needs for backtests.

• Fama-French exposures (factor_regression):
  • Model: daily excess returns regressed on Mkt-RF, SMB and HML (Fama & French 1993), optionally plus momentum (Carhart 1997). Window: 3 years, with Newey-West t-statistics.
  • Variance shares: each factor's share of return variance is β·cov(factor, return)/var(return). The shares sum to R², and the remainder is stock-specific.
  • Return attribution: the risk-free rate, plus alpha, plus β × each factor's average premium.
  • Readout: each stock gets a dominant driver and plain-English tilts (large-cap, growth/value, aggressive/defensive, significant alpha). rolling_exposures charts how betas drift over time.
• Strategies (run_strategies). Every strategy is rank-based on the watchlist, trades on the next day's returns (no look-ahead), and is charged 5 bps per unit of turnover:
  • Momentum 12-1, Jegadeesh & Titman (1993): long the top 3 / short the bottom 3 by 12-month return excluding the latest month, plus a long-only version. Monthly rebalance.
  • Short-term reversal: 1-week (Lehmann 1990) and 1-month (Jegadeesh 1990). Long recent losers, short recent winners.
  • Every strategy is then regressed on the Carhart factors, to see whether it earns alpha or just repackages known factors.
  • strategy_signals gives today's momentum rank and leg, the 5-day reversal rank, and a Bollinger z-score (oversold/overbought).
• Composite score (factor_scores):
  • Value: earnings yield, FCF yield, book-to-market and sales yield (Basu 1977; Fama & French 1992).
  • Quality: ROE, cash ROA, leverage and accruals (Sloan 1996; Novy-Marx 2013; Asness, Frazzini & Pedersen 2019).
  • Momentum: 12-1 and 6-1 returns.
  • Scoring: each metric is ranked across the watchlist, because ranks resist outliers when there are only 10 stocks. Ranks are averaged into three 0–100 pillars, then combined with equal weights. For banks, the cash-flow, sales and leverage ratios are skipped.
  • Point in time: trailing-twelve-month figures come from 10-Q/10-K filings, with Q4 = annual − Q1–Q3 and year-to-date cash flows differenced. A filing counts only from its SEC filing date. Share counts are converted to today's split basis so market cap matches split-adjusted prices (NVDA, AAPL, TSLA, GOOGL and AMZN all split during the decade).
• Evaluation (--evaluate): the Spearman rank IC of each pillar against the next 21-day return, over about 106 monthly cross-sections.

Read the results with two caveats. First, the watchlist is today's mega-caps, chosen with hindsight, so backtests are survivorship-biased. Second, ranking within 10 stocks is very noisy. On current data no pillar predicts next-month returns (every |t| < 2), and 12-1 momentum adds return only in its long-only form. Treat the score as a structured summary of value, quality and trend, not as a proven edge. Phase 6 re-runs these tests on point-in-time S&P 500 members: the momentum alpha disappears there, and the score still shows no predictive power.

Phase 4: autonomous paper-trading bot

python -m quant.trading_bot signals             # today's BUY / SELL / HOLD with rationales (no trading)
python -m quant.trading_bot run                 # trade the latest session(s) in the "paper" portfolio
python -m quant.trading_bot status              # positions, P&L, stops, win rate, recent trades
python -m quant.trading_bot backtest --save     # walk-forward test since 2018 (also seeds Kelly stats)
python -m quant.scheduler                       # run everything automatically (APScheduler)

Signals (make_signal). Four components in [−1, 1] are blended into one score:

Component	Weight	Built from
Trend	35%	50/200-day MA alignment (golden/death cross, price vs 200-day) 60%, plus the 12-1 momentum rank 40%
Fundamentals	25%	SEC quality pillar 60%, plus latest-quarter net income growth year-on-year 40%. A TTM loss or negative operating cash flow caps it at −0.5
Valuation	20%	Fama-French-style value pillar (E/P, FCF/P, B/M, S/P) 75%, plus the FF3 alpha t-stat 25%
Sentiment	20%	FinBERT score of on-topic headlines over the last 7 days, half-life 2 days, at least 2 headlines

• BUY: score ≥ 0.25 and an up-trend, unless the market regime is Bear.
• SELL: a held stock's score falls to ≤ −0.10, or its trend breaks.
• Every signal is stored in signals with a rationale that leads with the reasons pointing the way of the action. For example: "Buy XOM (score +0.53): 50 MA 7.9% above 200 MA, …; value 89/100 (E/P 4.4%) …; bullish news sentiment (FinBERT +0.33, 29 headlines); positive SEC earnings report (10-Q for 2026-06-30: net income +105% y/y) …".

Risk management:

• Sizing: volatility parity (weights ∝ 1/60-day vol). Alternatively --sizing kelly, where half Kelly (f = W − (1−W)/R, from closed round trips) sets the weight of a median-volatility position, scaled by vol. Until the live ledger has 20 closed trades, Kelly is bootstrapped from the saved backtest.
• Limits: at most 6 names and 20% per name. Gross exposure is 100% / 70% / 35% in Risk-On / Neutral / Risk-Off. Positions are resized only when 3+ points of equity off target.
• Stops: the initial stop is 3.5×ATR below entry, never more than 15% below it. A trailing stop follows at the highest close − 3.5×ATR and only moves up. A stop is checked against the session low and fills at the stop, or at the open after a gap down.

Ledger (SQLite or Postgres):

• portfolios holds cash, last session processed and parameters.
• positions holds holdings with their stop state.
• trades logs every fill with its reason (entry/add/trim/exit/stop), a plain-English rationale, the full signal JSON and round-trip P&L.
• portfolio_snapshots stores the daily equity curve.
• signals stores every day's signals.
• Runs are idempotent. After downtime, stops are replayed for every missed session, and new trades happen only on the latest one. Splits are applied to open positions.

Conventions: market-on-close fills with 5 bps slippage, whole shares, long only, no leverage. News counts only up to the close.

Automation (quant/scheduler.py, US/Eastern):

• 17:30 Monday–Friday: ingest everything, then run the bot.
• Every 2 hours, 08:00–20:00: refresh the news.
• Saturday 10:00: refresh the backtest.
• Logs go to data/logs/scheduler.log. The scheduler only runs while the process is up. On a laptop, Windows Task Scheduler is more reliable.

Windows scheduled task (registered on this machine as QuantPlatform daily cycle):

• When: Tuesday–Saturday at 07:30 KST. Korea has no daylight saving but New York does, so that is 17:30 New York in winter (EST) and 18:30 in summer (EDT). Either way it runs after the close.
• Runs: .venv\Scripts\pythonw.exe -m quant.scheduler --run-now in the project folder. pythonw means no console window; printed output goes to data/logs/scheduler.out.
• Missed runs: if the PC is off or asleep at 07:30, the run starts as soon as it is available. Market holidays are no-ops.
• Failures: the task's Last Run Result is 1 and the traceback is in scheduler.log.

Get-ScheduledTaskInfo -TaskName "QuantPlatform daily cycle"     # last / next run, last result
Start-ScheduledTask -TaskName "QuantPlatform daily cycle"       # run now
Unregister-ScheduledTask -TaskName "QuantPlatform daily cycle"  # remove

Backtest (2018–2026, 10 stocks, signals weekly, stops daily):

Sizing	CAGR	Vol	Sharpe	Max drawdown	Avg invested	Win rate	Payoff
Volatility parity (default)	8.7%	14.3%	0.46	−24.8%	66%	44%	1.86
Half-Kelly (full Kelly ≈ 14%)	6.0%	8.8%	0.39	−20.1%	37%	44%	1.94
SPY buy-and-hold	14.4%				100%

The bot is a risk-reduced trend follower. It cuts drawdowns but lags buy-and-hold in this bull-market decade, and it shows no edge over SPY, even on a survivorship-biased universe. On point-in-time S&P 500 members (Phase 6) it loses money: −1.8% a year.

• Stop width (2.5–5×ATR) barely changed returns (8.7–9.9% CAGR), but tight stops multiplied turnover; see the notes in trading_bot.BotParams.
• Exits: nearly every exit comes from the trailing stop, so the score effectively governs entries.
• Sentiment: historical sentiment is neutral, because the news archive only starts in October 2026.

This is a research tool, not investment advice.

Phase 6: point-in-time research universe (S&P 500)

Phases 3 and 4 were tested on today's 10 mega-caps, picked with hindsight. That universe is survivorship-biased, and ranks over 10 names are mostly noise. Phase 6 adds an optional research universe: every stock that was in the S&P 500 at some point in the price history, each one eligible only on the days it was actually in the index. The watchlist still drives news, live signals, the paper portfolio and the dashboard.

python -m quant.ingestion.pipeline --universe sp500 --steps universe prices fundamentals  # first run ~20 min
python -m quant.quant_models --evaluate --universe sp500           # deciles, point-in-time membership
python -m quant.quant_models --evaluate --universe sp500 --no-pit  # today's members back-filled (biased on purpose)
python -m quant.trading_bot backtest --universe sp500              # the Phase 4 rules on the same universe (~1 h)
python scripts/calibrate_phase2.py anomalies --universe sp500      # anomaly rules, member sessions only (~50 min)

To make the S&P 500 the default for these commands and for the daily pipeline, set RESEARCH_UNIVERSE=sp500 in .env. To extend headlines and FinBERT to the research universe as well, set NEWS_UNIVERSE=research; it is slow, so it's off by default.

Membership (quant/ingestion/index_membership.py, table index_members):

• Sources: Wikipedia's "List of S&P 500 companies" (today's members and the date each joined) and "Historical components of the S&P 500" (dated additions and removals since 1976). Until August 2026 the latter was the "Selected changes" table on the first page. Both pages are cached for a day in data/cache/.
• Reconstruction: the change log is replayed backwards from today's list. Undoing an addition means the stock wasn't a member before that date; undoing a removal means it was. Each stint in the index is one row (ticker, name, added, removed): added is inclusive, removed exclusive, and NULL means "before the log starts" or "still a member". A change takes effect before the open, so a stock removed on day D is no longer a member on D.
• Tickers: stored in Yahoo's format (BRK.B → BRK-B) and under the symbol the company trades as today, because Yahoo files a renamed company's whole history under its new symbol (Facebook's 2016 prices live under META). RENAMES lists 44 dated symbol changes from 2016 to 2026, so reused symbols resolve correctly ("Q" was QuintilesIMS until 2017 and is Qnity today). Rows that only record a symbol change (WLTW → WTW) are not counted as membership changes.
• Checks: the rebuilt index holds 489–504 names at every month-end since 2010. No row in the log contradicts the reconstruction after March 2015. The remaining conflicts (Allergan's symbol passing to Actavis in 2015; Delphi, Joy Global and Ingersoll-Rand in 2010–2012) predate the price history. A name check against SEC's current registrants found no symbol whose Yahoo history overlaps the stint of a different company.

Ingestion (--universe sp500):

• Prices: market_data.update_daily_batch downloads daily bars and split events for 100 tickers per Yahoo request: 682 tickers and 1.44 M bars in about 6 minutes. Later runs fetch only the last week and stop asking for delisted names. Intraday bars stay watchlist-only. Tickers Yahoo has no data for are remembered in data/cache/yahoo_no_data.json and re-tried monthly.
• Fundamentals: the same SEC client, at ≤ 6.7 requests/second (SEC allows 10): 606 filers and 24,105 filings in about 10.5 minutes, then only new filings. Share classes of one filer (GOOG/GOOGL, FOX/FOXA, NWS/NWSA) are fetched once; the others reuse those filings through their shared CIK.
• News: headlines and FinBERT stay watchlist-only unless NEWS_UNIVERSE=research.

Point in time in the models:

• Universe.eligible(date): index members on that date with a price that day. A plain watchlist (members=None) keeps every ticker eligible, so watchlist results are unchanged; they were checked byte for byte against the pre-change outputs.
• Scores (factor_scores, score_history, ICs) rank only the stocks eligible on each date.
• Strategies (run_strategies) mask non-members' signals. Each leg holds a decile of the stocks ranked that day (--quantile) instead of a fixed 3 names. The equal-weight benchmark holds the previous session's members.
• Bot (trading_bot backtest): signals and entries come only from members. A holding that leaves the index is sold at the last close. Live trading stays on the watchlist, and --save stores the run as backtest_sp500, so the backtest portfolio used for the Kelly bootstrap is not touched.

Results: 10 hand-picked stocks vs point-in-time S&P 500

Same prices, same 2016–2026 window, net of 5 bps per unit of turnover. The middle column (--no-pit) ranks today's 503 members over the whole history: the usual survivorship-biased setup, kept only to measure the bias.

Strategy: CAGR, Carhart alpha (t)	10 mega-caps, top/bottom 3 (before)	Today's S&P 500, back-filled, deciles	Point-in-time S&P 500, deciles
Momentum 12-1, long only	37.0%, +14.8% (2.6)	28.3%, +8.8% (2.8)	14.4%, −2.2% (−0.8)
Momentum 12-1, long/short	1.4%, −1.2% (−0.1)	2.2%, −1.3% (−0.3)	−3.7%, −5.8% (−1.4)
Reversal 1-week, long/short	−10.5%, −8.9% (−0.9)	−1.6%, −2.8% (−0.4)	−0.8%, −2.3% (−0.3)
Reversal 1-month, long/short	−8.7%, −4.5% (−0.5)	−6.8%, −4.9% (−0.8)	−2.0%, −0.4% (−0.1)
Equal-weight universe	28.9%, +10.3% (3.8)	17.1%, +3.6% (2.6)	12.1%, −0.5% (−0.4)

Score IC vs next 21-day return (t)	10 mega-caps	Today's S&P 500, back-filled	Point-in-time S&P 500
Stocks per monthly cross-section	10	472–503	434–503
Value	−0.056 (−1.27)	−0.003 (−0.14)	+0.001 (+0.04)
Quality	+0.033 (+0.86)	+0.018 (+1.55)	+0.017 (+1.46)
Momentum	−0.001 (−0.02)	−0.002 (−0.11)	−0.006 (−0.29)
Composite	−0.026 (−0.75)	+0.006 (+0.46)	+0.006 (+0.43)

What changed:

• The momentum edge was survivorship. The 10-stock momentum result was mostly hindsight: knowing which companies would become mega-caps. Back-filling today's S&P 500 still shows +8.8% alpha (t 2.8). With point-in-time membership it disappears (−2.2%, t −0.8), and long-only momentum roughly matches the equal-weight index (14.4% vs 12.1%, at higher volatility). Even the equal-weight benchmark loses its "alpha" (+3.6% → −0.5%) once the index is not picked with hindsight.
• The composite score still predicts nothing, now with real breadth. About 480 names per month make the IC far more precise than with 10 names: a standard error of about 0.014 against about 0.035 before. Quality comes closest (t 1.46), still below the usual bar of 2. Momentum's IC is about zero, which is no surprise: in the stored factor data, Fama-French momentum earned 0.1% a year in 2009–2026, against 9.7% in 2000–2008 (4.0% over this window).
• Short-term reversal is no longer a big loser. On 10 names its −9% to −11% was concentration noise. Across the index, one-month reversal is about flat (−2.0%, alpha t −0.1). One-week reversal turns over 175× a year, so 5 bps per trade costs about 9% a year.

Bot backtest (Phase 4 rules unchanged, volatility parity, 2018-01-03 to 2026-10-19; about an hour on the S&P 500, since it scores ~500 stocks every week):

Universe	CAGR	Vol	Sharpe	Max drawdown	Round trips	Win rate	Payoff	Full Kelly
10 mega-caps (before)	8.7%	14.3%	0.46	−24.8%	292	44%	1.86	14.2%
Point-in-time S&P 500	−1.8%	15.6%	−0.22	−49.1%	385	35%	1.62	0%
SPY buy-and-hold	14.4%

The bot's modest result on 10 names came from trading only stocks already known to be winners. Picking 6 names from ~500 by a score with no predictive power, and exiting almost entirely on trailing stops (384 of 385 exits), loses about 0.5% per round trip, and full Kelly is 0%, so Kelly sizing would invest nothing. Most of the damage came in 2020 (−24.7%) and 2023 (−11.7%); 2024–2026 were positive (+8.5%, +12.9%, +13.3%).

In the COVID crash the bot bought a fresh set of six names on three weekly signal days in a row (26 Feb, 4 Mar and 11 Mar 2020). The 50/200-day trend filter lags, and new buys are blocked only once the regime turns Bear. Each set was stopped out within days, realizing −$21.7k on a $100k book in a month. The bot was then mostly Risk-Off and only 37% invested during the rebound (+3.2% from March to August). Tightening entries in fast drawdowns (blocking buys while the regime is Volatile, or after a market stop-out) is the obvious next experiment, but it should be judged on this universe, not on the 10 names. A trade-by-trade check found no data artifacts: the worst round trip is −20.5% (Humana gapping below its stop in March 2020), and no fill is at an implausible price.

Phase 2 anomaly rule (calibrate_phase2.py anomalies --universe sp500). The run covers 560 stocks with at least a year of member sessions (4,629 member-years, 19,698 earnings releases from 8-K Item 2.02 filings). The tuned daily rule holds up, so it was not changed:

Daily rule (abs / z / min)	Universe	Flags per stock-year	Spread across stocks	Earnings reactions caught	Flags on earnings days	Market-driven flags
5% / 3σ / 2% (old)	10 mega-caps	13.8	5.8× busiest/calmest	43%	15%	33%
8% / 3σ / 2% (tuned)	10 mega-caps	8.7	2.9× busiest/calmest	35%	19%	32%
5% / 3σ / 2% (old)	point-in-time S&P 500	12.6	2.9× p90/p10	40%	15%	24%
8% / 3σ / 2% (tuned)	point-in-time S&P 500	8.8	1.7× p90/p10	37%	18%	25%

Across the index, a single busiest/calmest ratio is dominated by one outlier, so the spread is shown as p90/p10. A reaction day moves 3.9× a normal day for the median stock (1.1× the day before, 1.5× the day after), so the 8-K timing holds beyond mega-caps. The intraday rules use 15-minute bars, which exist only for the watchlist, so they are unchanged.

Limitations

• Missing prices: 7.4% of member-days since October 2016 can't be ranked. Yahoo deletes the history of most companies that stop trading:

  • Acquired or merged: 77 companies, 6.0% of member-days (Twitter, Activision, Celgene, Xilinx, Discover, Juniper, Ansys, Hologic and others). Warner Bros. Discovery, AvalonBay and Electronic Arts, all acquired in 2026, survive only as a one-day stub. Takeovers usually end with a premium, so this gap leaves out gains.
  • Removed from the index, then delisted: 16 names, 1.1%. Chesapeake, Frontier and Mallinckrodt went bankrupt after leaving; Dish, Comerica, Nordstrom and others were acquired.
  • Failed while in the index: SVB, Signature Bank and First Republic, 0.2%. Their collapses are missing, which flatters long positions.
  • Other, 0.2%: the pre-2019 stints of 21st Century Fox class B and Dow Chemical, whose symbols now belong to their successors, plus the first three sessions of the window (the batch download started on 20 Oct 2016, the watchlist on 17 Oct).

  A stock without data is never ranked, so the screen still knows, with hindsight, which companies will vanish. Fixing this needs a delisting-aware price source such as CRSP or Norgate, and neither is free.

• Membership comes from Wikipedia. It is volunteer-maintained and cited to S&P press releases, but it is not S&P's official file. The symbol-change list (RENAMES) was compiled by hand. Effective dates are used, while index changes are announced about a week earlier.

• History is short: about 10 years of prices (2016–2026) and 44 filings per company. The Fama-French factors end on 2026-08-31, so alphas stop there.

• Share classes double-count: GOOG/GOOGL, FOX/FOXA and NWS/NWSA rank as two names with nearly identical scores, as they do in the index itself.

• Splits: for the research tickers, split history covers only the 10-year download window. A split between a stock's earliest filings (late 2015) and the start of the window (October 2016) would leave those filings' share counts on the wrong basis.

• Same caveats as before: 5 bps per unit of turnover, no borrow costs on shorts, and no historical news sentiment for the bot.

Phase 5: web dashboard

streamlit run dashboard/app.py          # http://localhost:8501

There are four pages, all reading the same database the pipeline and the bot update. The heavy point-in-time context is loaded once per process and cached for 15 minutes. Use Reload data in the sidebar after a pipeline run. The sidebar also shows how fresh each source is.

Page	What it shows
Market Overview	Regime, stance and risk score with the reasons behind them; the S&P 500 with its 50/200-day MAs and Risk-On / Neutral / Risk-Off shading; ten FRED indicator tiles with 2-year sparklines; watchlist movers with factor scores; breaking news with FinBERT tone and links
Stock Deep Dive	Daily or 15-minute candlesticks with the 50/200-day MAs, optional Bollinger bands, and separate volume and RSI panels; ▲/▼/● news markers on the bar each headline first affected; ★ unusual moves whose hover shows the Phase 2 explanation; tables of moves (with catalyst links) and headlines (with the 1-hour price reaction)
Financials & Research	TTM revenue, net income, margins, FCF, EPS and D/E from SEC XBRL; quarterly revenue, net income and margins (Q4 derived); the filings list with SEC links; Fama-French / Carhart exposures with t-stats; a return-attribution waterfall; rolling 1-year betas; the watchlist exposure table; composite factor scores; academic strategy backtests
Live Trading Bot	Paper or backtest portfolio: equity, return vs SPY, cash, win rate and payoff; equity vs SPY indexed to 100 with drawdown; open positions with weights and distance to stop; the latest signals as component-contribution bars with full rationales; the filterable trade ledger (CSV download); the distribution of closed-trade returns by exit reason

Charts follow one validated palette. It was checked with a colorblind-safety validator in light and dark mode, and dark mode uses its own steps rather than an inversion.

• Color: candles are blue up / red down (not red/green). Regime shading uses status colors with labels.
• Axes: every chart has a single y-axis. Price, volume and RSI are separate panels, and the bot is compared with SPY indexed to 100.
• Data: every chart has a table or hover twin.

tests/test_dashboard.py renders every page headlessly with Streamlit's AppTest (it is skipped until the pipeline has filled the database).

Global universe: stocks, ETFs and crypto worldwide

python -m quant.universe.refresh                         # discover, prices, fundamentals, score (~30 min daily)
python -m quant.universe.refresh --steps score           # re-score only
python -m quant.universe.refresh --fund-limit 4000       # fetch more fundamentals in one run

The dashboard's Global Overview page (now the start page) screens the whole investable universe with the same processes the deep tier uses. You can then promote picks to your watchlist for the full treatment.

Two tiers:

	Broad tier (Global Overview)	Deep tier (watchlist)
Coverage	~8,700 stocks ≥ $1B in 21 markets, ~2–3k liquid US-listed ETFs, top 250 coins	The default 10 + anything you promote
Prices	2 years daily, Parquet (data/universe/)	10 years daily + 15-minute bars, database
Analysis	Trend (50/200 MA, crosses), momentum, risk, unusual moves, value, quality, analyst consensus, market regime per region; BUY / HOLD / AVOID with a rationale	Plus news with FinBERT sentiment and the relevance filter, explained moves, SEC filings (US), Fama-French, the paper-trading bot (USD-quoted names)

Sources (all free, no keys):

• Yahoo Finance's screener supplies 250 quotes per request. It covers market cap, 50/200-day averages, current-year P/E, dividend yield and analyst rating.
• Yahoo prices are downloaded in batches. Per-company ratios (including gross, operating and net margins) come from Yahoo's quote summary, refreshed for about 1,500 names per run.
• CoinGecko supplies the coin list. Nasdaq Trader's directory lists every US-listed ETF.
• Scraping sites like TradingView or Investing.com breaks their terms and breaks whenever their pages change, so they aren't used.

How instruments are selected:

• Markets are the US, Canada, the UK and Europe (DE, FR, NL, CH, IT, ES, SE, DK, FI, NO, BE), Japan, Korea, China, Hong Kong, Taiwan, India and Australia. The $1B floor is converted to each market's currency. UK prices are quoted in pence, but Yahoo reports UK market caps in pounds.
• Cross-listings (an ADR plus its home listing, dual share classes, foreign lines on Xetra or the LSE) are de-duplicated by company name, keeping the most liquid listing. The others are shown as "also listed".
• Crypto excludes stablecoins (including EUR and JPY coins), wrapped, staked or bridged copies, and tokenized funds, T-bills and private credit.
• Each coin's Yahoo ticker is checked against CoinGecko's price. Yahoo suffixes tickers that several coins share (Uniswap is UNI7083-USD), and a plain ticker can price a different coin (ARB-USD is not Arbitrum). Coins Yahoo doesn't carry are listed without technical scores. Results are cached in data/universe/crypto_symbols.json.
• ETFs must trade at least $1M a day on average.

How instruments are scored (quant/universe/scoring.py):

• Pillars: trend, momentum (12-1 and 6-1, on calendar days so crypto compares fairly), risk (volatility, 1-year drawdown), value (forward and trailing E/P, B/M, EBITDA/EV, dividend yield), quality (ROE, margin, low debt/equity, growth) and analysts. ETFs also get cost (expense ratio).
• Peers: each pillar is a percentile among peers. For stocks that means their market (and market × sector for value and quality); ETFs and coins rank within their class. The composite is the mean of the available pillars.
• Signals: BUY candidate = up-trend + composite ≥ 60 + the market isn't in a Bear regime. AVOID = down-trend + composite ≤ 40.
• Market regimes: each market's regime comes from its main index: S&P 500, TSX, STOXX 600, Nikkei 225, KOSPI, Hang Seng, TAIEX, Nifty 50, ASX 200, and Bitcoin for crypto. The US uses the VIX; elsewhere the index's realised volatility.
• Unusual moves: the Phase 2 rule, applied to every instrument every day.

Watchlist → deep tier:

• In Global Overview, select rows and click Add to watchlist (picked up by the next daily run), or Add & analyse now, which immediately fetches prices, SEC filings and news for stocks and ETFs.
• Non-US names get news by company name ("Samsung Electronics"), with that name as a relevance alias.
• The paper bot's ledger is in dollars, so it only trades USD-quoted names. Crypto stays in the broad tier.
• The scheduled daily cycle covers the deep tier, runs the bot, and then refreshes the global universe (--no-universe skips that).

Limits:

• Stock data is Yahoo's, so it's ~15 minutes delayed and corporate data can lag.
• Per-company ratios cover the universe only after about a week of daily runs. Yahoo rate-limits the per-company calls, and the fetch stops cleanly and resumes next run. Until then, value and quality scores rest on fewer inputs.
• Yahoo sometimes publishes a market's latest session late (for example, Monday's European bars on Tuesday morning in Asia). Those names score on the previous close, and the next refresh fills the gap.
• Analyst consensus stands in for news sentiment in the broad tier.
• Crypto in the broad tier updates on scheduled weekdays only.
• Rankings are relative to peers, not predictions; see the Phase 3 evaluation.

Live website (GitHub Pages)

https://chanwoo777.github.io/quant-platform/ is a static copy of the dashboard, refreshed once a day, plus tools that run in the browser. GitHub Pages can't run Python, so the PC that runs the pipeline exports the data as JSON and pushes it.

python -m quant.web.export                     # build the site in data/site (~1 min)
python -m quant.web.publish                    # export + push to the gh-pages branch
python -m quant.scheduler --run-now --publish-site   # daily cycle, then publish

Page	What it does
Global	The screener (every stock ≥ $1B, liquid US ETF and top coin) with filters, per-symbol charts and signal reasons, plus a sector heatmap of the day's moves. ★ adds a name to your watchlist.
Compare (#/compare/<A,B,C>)	2-5 tickers side by side: returns, risk, factor exposures, fundamentals and a correlation heatmap. Details below.
Markets	Regime and stance, the S&P 500 stance chart, FRED tiles and news. Your watchlist: search to add any stock, ETF, coin or index, ✕ to remove. It is saved in the browser (localStorage).
Stock (#/stock/<ticker>)	Any of ~11,400 symbols: candlesticks with every indicator below, plus returns, risk and peer percentiles. Deep-tier names add news markers and explained moves. The chart loads about two years of daily bars in the browser (see Price history below).
Financials (#/financials/<ticker>)	A research view for any ticker: fundamentals, the Analysts pillar (with Finnhub's recommendation trends when a key is set), a peer table, financial statements and the factor model. Statements come from SEC filings for the deep tier, or Finnhub's reported statements for US names when a key is set.
Quantum (#/quantum)	A beginner's guide to quantum algorithms for investors: basics, a qubit lab, option pricing, VaR/CVaR and portfolio labs with classical baselines, a readiness scorecard and the nightly simulator run. Details below.
Live (#/live/<symbol>)	The Pro chart: a TradingView-style chart drawn by the site, with every indicator and setting under your control, and price alerts. Details below.
Methodology (#/methodology/<id>)	How every score, signal and metric is calculated, from the code; every "?" on the dashboard links here.

The header search autocompletes over the whole universe. Pretty links such as /quant-platform/stock/AAPL also work: 404.html turns them into the hash route.

Factor models in the browser. docs/assets/quant.js reimplements quant_models.factor_regression and rolling_exposures: OLS with Newey-West (5-lag) errors, variance shares, return attribution and tilts. So FF3, Carhart and FF5 run for any stock, from factors.json and the daily prices the browser loads. The loadings chart shows 95% confidence intervals. tests/test_site_js.py runs the JavaScript under Node.js against the Python models, on synthetic data with missing values; they agree to about one part in a million. Fixes in this round:

• Rolling betas: the site's rolling-beta chart sampled an already monthly series a second time, leaving about 22 points over 9 years. It now keeps every monthly point.
• FF5: Fama & French (2015) is supported end to end. Ingestion stores the five-factor file's RMW, CMA and size factor (smb5, built from more sorts than SMB). quant_models.FF5 and the Streamlit Financials page offer it, and the site estimates it per ticker.

Indicators. These are the popular TradingView indicators, implemented in quant.js from the Pine Script built-ins:

• Overlays: SMA 50/200, EMA 9/21, Bollinger Bands (20, 2), VWAP (session), Ichimoku Cloud, Supertrend (10, 3) and Parabolic SAR.
• Oscillators: volume, RSI 14, MACD (12, 26, 9), Stoch RSI, Stochastic (14, 3, 3), ATR 14, ADX/DMI 14 and OBV.

They are drawn on the Stock pages and the Markets S&P 500 chart. Your choices are remembered in the browser. tests/test_site_js.py checks each one under Node.js against Python versions of the Pine definitions, and against the ta library where it uses the same formulas. Every chart has highs, lows and volume, because prices load in the browser as full daily bars (indices have no volume).

Pro chart (Live). A TradingView-style chart built on TradingView's open-source Lightweight Charts v5, so indicators aren't limited the way an embedded widget is.

• Toolbar: symbol search, intervals from 1 minute to 1 month, and chart types (candles, hollow candles, bars, line, area, Heikin Ashi). Also log and percentage scales, screenshot and full screen.

• Indicators: a searchable dialog with 26 studies, each with TradingView's default inputs:

  • Trend: moving average (SMA, EMA, WMA, HMA, VWMA or RMA), Donchian, Ichimoku, Supertrend, Parabolic SAR, classic pivots, 52-week high/low and ADX/DMI.
  • Momentum: RSI, MACD, Stochastic, Stoch RSI, Williams %R, CCI, ROC and relative strength vs a benchmark (O'Neil RS line or Mansfield).
  • Volatility: Bollinger Bands, %B, BandWidth, Keltner Channels and ATR.
  • Volume: volume, VWAP, anchored VWAP, MFI and OBV.

  Add any study as often as you like. Each opens with editable inputs and colours (⚙ in the legend), can be hidden or removed, and oscillators get their own resizable pane.

• Legend: OHLC and every indicator value under the cursor, as on TradingView. It folds away on phones.

• Drawing tools: trend line, ray, horizontal and vertical lines, Fibonacci retracement, rectangle and a measure tool. Drawings are saved per symbol in the browser.

• Shortcuts: TradingView's own, such as / for indicators and typing 5, 60, 1D or W then Enter for the interval. Also Alt+T, H, V and F for drawing tools, Alt+L for log scale, Alt+S for a screenshot, Esc, Delete and Ctrl+Z.

• Investor playbooks: the indicator sets well-known traders are known for, with their published settings and usual interval:

  • Paul Tudor Jones (200-day line)
  • Mark Minervini (Trend Template)
  • William O'Neil (CAN SLIM)
  • Stan Weinstein (stage analysis, weekly)
  • the Turtles (Donchian 20/55, ATR)
  • John Bollinger
  • Welles Wilder
  • Linda Raschke (Holy Grail, 3-10)
  • Larry Connors (RSI 2)
  • Gerald Appel (MACD)
  • Goichi Hosoda (Ichimoku)
  • Brian Shannon (VWAP)
  • Larry Williams (%R)
  • Joseph Granville (OBV)

  Fundamental investors like Buffett and Lynch don't use chart indicators, and the page says so rather than inventing a set.

• Data: every source is keyless and called from the visitor's browser:

  • Crypto: Binance public market data (REST history + WebSocket stream).

Research tools. Five tools added on top of the pages above. Each uses only the nightly export and the in-browser price history (see Price history); no new data sources.

• Compare (#/compare/AAPL,MSFT,NVDA): 2-5 tickers side by side. The tickers live in the URL, so a link shares the comparison. The table has a column per ticker, best/worst highlighting per row and a CSV download.

  • Returns: 1M, 3M, YTD, 1Y and 3Y annualised, plus a growth-of-100 chart.
  • Risk: volatility, max drawdown, Sharpe and Sortino, and beta vs the S&P 500, over a 1, 2 or 3-year window. The risk-free rate is FRED DGS3MO, daily.
  • Factors: exposures to the chosen Fama-French model, with t-statistics and R². The 3-year window uses weekly prices and weekly-compounded factors.
  • Fundamentals: each cell is labelled with its source, SEC EDGAR for the deep tier and Yahoo otherwise. Gross and operating margins are Yahoo's trailing twelve months for every name; a name the rolling refresh hasn't reached since they were added shows n/a.

• Correlation heatmap: Pearson correlation of daily returns (3M, 6M, 1Y) or weekly returns (3Y). It can be ordered by average-linkage clustering on 1 − correlation, or alphabetically, and shows the average pairwise correlation and the most and least correlated pairs. One component (drawCorrelationPanel) serves both Compare and the Quantum portfolio lab.

• Sector heatmap (Global → Sector heatmap): a squarified treemap of the latest nightly close's moves by sector. Tile size is market cap; colour is the cap-weighted 1-day change. It drills down to each sector's largest 150 stocks. The sector is the exported Yahoo sector field, which every stock has, so no SIC mapping is needed. Missing values show as grey n/a tiles.

• Price alerts (Live): price at or above or below a level, or a % move from the previous close. They are saved in the browser and checked every 15 s while the tab is open, on any page:

  • Stocks, ETFs and indices: one batched CNBC quote request. Markets without a quote, such as Korea, use CNBC's 1-day chart and can't check % alerts.
  • Coins: Binance's trading-day ticker.

  Each alert fires once, as a browser notification, or an in-page toast if permission is denied. It can then be re-armed or deleted.

• Methodology (#/methodology/<id>): 43 entries written from the code, each with a description, the formula (MathML), inputs and sources, window, update frequency, limitations, and the file and function. A "?" next to each dashboard metric links to its entry.

tests/test_research_js.py checks the maths under Node.js against numpy:

• Sharpe, Sortino, beta, volatility, factor betas and R² against a hand calculation;
• the correlation matrix, symmetric with 1 on the diagonal;
• clustering, the treemap layout and grouping with missing fields;
• the alert conditions.

Quantum tab (#/quantum). A beginner's guide to quantum algorithms and where they might help investors, with a classical baseline next to every quantum number. Twelve sections:

1. Basics: superposition, measurement, entanglement and interference, the "tries every answer at once" myth, and why speedups need structure.
2. Qubit lab: a one-qubit widget (H, X, Z), a 1-5 qubit circuit simulator (H, X, Z, CNOT) with amplitudes and 1,000 measured shots, and an animated Grover search over 8 or 16 items. All of it runs in the browser.
3. Toolbox: eight building blocks, their speedups and finance uses.
4. Monte Carlo vs QAE: error against samples on a log-log chart (1/sqrt(N) vs 1/N), a target-error slider (0.1% means about 1,000,000 samples vs 1,000 queries), Monte Carlo measured in the browser, and the nightly QAE runs.
5. Option pricing: any stock in the universe. Spot and volatility (3, 6 or 12 months of daily log returns) come from the site's prices; the rate is FRED's 3-month T-bill yield (DGS3MO, now in market.json). The page shows Black-Scholes, Monte Carlo with a 95% interval and its convergence, and the nightly amplitude estimate with its error budget. A six-step explainer covers how QAE prices the option.
6. Risk: 1- and 10-day VaR and CVaR at 95% and 99% for a basket of up to 10 stocks, by historical simulation and Monte Carlo. The nightly QAE estimate of one basket's tail probability, and VaR found by bisection.
7. Portfolios: a long-only efficient frontier for 3-15 stocks, and the "hold K stocks" QUBO with its matrix, an exact brute force, and the nightly QAOA result compared with the optimum.
8. Readiness scorecard: a seven-step check per use case, ending in Promising / Open question / Not yet useful.
9. The nightly run: date, runtime, package versions, qubits, circuit depth and circuit diagrams.
10. Why it's not ready yet.
11. A security note for investors: Shor's algorithm and NIST's post-quantum standards.
12. Glossary, a 9-week study path and references.

The maths in docs/assets/quantum.js is tested under Node.js (tests/test_quantum_js.py):

• Black-Scholes against Hull's reference case and scipy;
• Monte Carlo converging at 1/sqrt(N);
• VaR, CVaR and the Monte Carlo VaR against numpy;
• the frontier against scipy's SLSQP;
• the QUBO and brute force against Python.

Nightly quantum simulations (quant/web/quantum.py). These use Qiskit's statevector simulator only: no quantum hardware, no accounts and no keys.

• Option pricing: iterative amplitude estimation of at-the-money European calls on AAPL, MSFT and NVDA at 30, 90 and 180 days. It uses a 5-qubit lognormal model and reports the price, a 95% interval, oracle queries and circuit depth. These sit next to Black-Scholes, Monte Carlo, the circuit's exact value and the exact value on the 32-price grid.
• Risk: the tail probability of a 5-stock basket at its historical 95% VaR, with the return histogram loaded into 5 qubits, and VaR by bisection.
• Portfolio choice: QAOA (2 layers, COBYLA) on the cardinality QUBO for 6 stocks, with K = 3, compared with brute force.

The export runs it in a subprocess with a 5-minute limit; it takes about 2.5 minutes here. The module replaces data/quantum.json only when every job succeeds. The export then copies the last good file into the site, so a failed or timed-out run leaves the previous night's results online. With no file at all, the page says the results aren't published yet, and every classical part still works. tests/test_quantum_job.py covers the failure paths and runs small real simulations.

pip install -r requirements-quantum.txt    # pinned: qiskit 2.6.1, qiskit-algorithms 0.4.0,
                                           # qiskit-finance 0.4.1, qiskit-optimization 0.7.0
python -m quant.web.quantum                # run the simulations alone (~2.5 min)
python -m quant.web.export --no-quantum    # export without re-running them

Limits. Every quantum number is labelled "Simulated on a classical computer - educational, not a trading signal".

• The circuits are tiny: 32 price points and 6 stocks. The deepest option circuit has about 190,000 layers, far beyond today's noisy hardware.
• At this size QAE needs about as many queries as Monte Carlo needs samples, and QAOA does no better than a random valid portfolio.
• The page says plainly that no practical quantum advantage on a real finance problem has been shown as of 2026.

Price history. The site publishes no price history, because Yahoo's terms allow personal use only. The nightly export publishes calculated numbers instead (scores, percentiles, returns, volatility, drawdowns, betas, signals), plus the latest close and fundamentals. Analyst ratings aren't published either: they are licensed data, so the site shows only the calculated Analysts pillar, and the signal reasons name that percentile instead of the rating. Every chart, the Financials factor model and the Quantum labs load past prices in the visitor's browser (window.LIVE.dailyHistory in live.js):

• Stocks, ETFs and indices: about two years of daily bars from CNBC's 1Y range, or ten years of weekly bars from its 5Y range. At most four requests run at once, because CNBC throttles bursts.
• Coins: daily bars from Binance.
• Not covered: Tokyo and Hong Kong listings, and coins without a Binance market, show their scores and returns without a chart.
• Adjustment: CNBC's prices may not be adjusted for dividends, which slightly lowers measured returns and alpha.

The nightly quantum simulations still use the platform's own price store on the PC; their inputs and results are published, not the prices.

Data sources and keys:

• No key needed: scores, fundamentals and analytics come from the nightly export. Prices load in the browser from CNBC and Binance's public market data (data-api.binance.vision, data-stream.binance.vision).
• Optional Finnhub key: entered in ⚙ on the site, it adds live US trades in Live, plus live quotes, recommendation trends and reported statements in Financials. It also enables lookups for US tickers outside the universe. The settings box labels it as a credential. The key stays in that visitor's browser, in session storage (forgotten when the tab closes) unless they tick "Remember on this device", and is sent only to finnhub.io. Nothing is stored in the repository, because anyone can read a key placed in a public site.
• Yahoo: Yahoo Finance has no browser API (no CORS), so symbols outside the universe can't be fetched from it live.

How it works:

• Code: the site's code is versioned on main under docs/:

  • index.html, about.html, privacy.html and 404.html
  • assets/app.js: the pages
  • assets/quant.js: statistics, factor models and indicators
  • assets/quantum.js: the Quantum page and its maths (state-vector simulator, option pricing, VaR, frontier, QUBO)
  • assets/live.js: the Pro chart (Lightweight Charts v5, Binance and Finnhub WebSockets, CNBC polling, studies, drawings, shortcuts), the in-browser price history and price alerts
  • assets/research.js: Compare and the shared correlation heatmap
  • assets/methodology.js: the Methodology page
  • assets/app.css

  It is plain JavaScript, with Plotly and Lightweight Charts from jsDelivr and no build step.

• Data: quant/web/export.py calls the same functions as the Streamlit pages (dashboard/common.py), so both show the same numbers. It writes JSON to data/site/data/:

  • meta, global (the screener), symbols (search) and indices
  • sym/<xx>: fundamentals and statistics per symbol (no prices), in 256 FNV-1a buckets
  • factors, market (with four years of daily DGS3MO for Compare's Sharpe), deep/<T>, research and quantum

  The total is about 11 MB raw; GitHub Pages compresses it.

• Publishing: quant/web/publish.py commits data/site as a fresh one-commit history and force-pushes it to gh-pages. Each day's snapshot replaces the last, so data never piles up in the repository. Pushing uses your normal git credentials (the GitHub CLI's credential helper here).

• Schedule: the Windows scheduled task runs --run-now --publish-site, so the site updates after each 07:30 KST cycle.

What differs from the local dashboard:

• The site is read-only. The watchlist lives in the visitor's browser. Promoting a name to the deep tier (news and SEC filings) still happens on the PC.
• News markers and explained moves cover the deep tier. Charts show about two years of daily bars (ten years weekly on the Markets chart), against the local dashboard's ten years.
• There are no 15-minute bars.

The site publishes no Yahoo price history and no analyst ratings. It still publishes the latest close and fundamentals from Yahoo Finance, whose terms allow personal use only; publishing those is the owner's decision. BUY / HOLD / AVOID are mechanical labels from ranking past data against peers, and the site says so next to them: they are not recommendations. To take the site down, disable Pages in the repository settings or delete the gh-pages branch.

Using the site with others

• Start here (#/start) is where a first visit lands. It has three one-click examples: Samsung against its big-tech peers on Compare, a basket in the VaR / CVaR lab, and an option priced classically and by the nightly quantum run.

• Shareable links. Every page writes what it shows after the route, for example #/global?sig=BUY&mkt=US&sort=r1m&page=2 or #/quantum/risk?syms=AAPL,TSM&w=50,50.

  • The 🔗 button copies that link, and it opens the same view in a fresh browser.
  • Links unfurl in KakaoTalk, Slack and Discord with a preview card (docs/assets/og-preview.png).

• Daily brief (Global). Each night quant/web/brief.py writes a few sentences by rule, not AI, in English and Korean. It covers:

  • the largest moves among stocks worth $10B or more;
  • the best and worst sector;
  • factor ETFs against SPY;
  • the VIX;
  • Treasury yields.

  The last 30 briefs are a click away.

• Watchlist. ☆ next to any ticker adds it. The list is kept in your browser and shown on Global and Live, and it can be exported and imported as JSON.

• PNG and CSV. Every chart downloads as a PNG marked with the site's address and the data's as-of date. Every table and chart downloads as CSV.

• Feedback. "Suggest a feature / report a bug" opens GitHub Issues, with forms for bugs, data errors and ideas (.github/ISSUE_TEMPLATE).

• 한국어 / English. The button in the header switches the language. The choice is remembered in the browser, and a Korean-language browser starts in Korean.

  • How it works: the pages are written in English. In Korean, docs/assets/i18n.js translates the interface as it draws, from the dictionary in docs/assets/i18n.ko.js, keyed by the English text.
  • Values: numbers, dates and tickers become placeholders. Dates and dollar amounts are written the Korean way (2026. 10. 27., 5.1조 달러).
  • Kept as is: tickers, company names, headlines and formulas.
  • Finding gaps: any interface text without an entry stays English and is listed by I18N.report() in the browser console, or on screen when ?i18n=missing is added to a page's address. That list is how to find hard-coded text.
  • Tests: checks that every entry keeps its placeholders and that alerts use the dictionary.

How to verify a number

Every number on the site can be traced to its source, its date and the code that computed it.

1. Check the freshness badge. Every panel heading carries a badge that names the panel's sources and the oldest as-of date among them.

   • Green: fresh. Amber: older than the source expects (one trading day for prices, about two months for the Fama-French library). Red: missing, failed, or more than three times too old.
   • Ages are counted in your browser, so a site that stops updating turns amber, then red, by itself.
   • Hover the badge for each source's date and age; click it for the Data status page.

2. Read the Data status page (#/status, linked from the footer). The nightly job writes data/status.json (quant/web/status.py). For each source it lists:

   • the last successful run and the date the data runs to;
   • row counts, missing tickers and stale series;
   • outliers flagged for review: daily moves beyond 5 standard deviations of the previous year;
   • the fail-soft quality checks (quant/web/quality.py): repeated rows, gaps of 3+ sessions, closes at or below zero, zero volume, and 20x one-day jumps (unit changes).

   A failing series is never dropped silently. Its symbol gets a ⚠ in the screener and a note on its pages. A missing file or a failed export step shows red, and the page that needs it says so instead of drawing a partial chart.

3. Look up the formula. The "?" next to a metric opens its Methodology entry (#/methodology/<id>). Each entry gives the formula, the inputs and their source, the window, the limitations, and the file and function that compute it.

4. Download the data. Every table and chart has a CSV button. Each file starts with the panel's title, its sources and as-of dates, the download time and page, and the disclaimer. The screener's file holds every matching row, not just the visible page.

5. Recompute it. The code is public ("View the code" in the header and footer). Run python -m quant.web.export to rebuild the same JSON from the same sources. The in-browser maths (docs/assets/quant.js) is tested against numpy in tests/test_research_js.py.

6. Check the signals' record. The Signal track record (#/track-record) shows forward returns of the BUY / HOLD / AVOID labels after 5, 20 and 60 trading days, against each market's index, with the hit rate, the number of signal dates and the start date.

   • How the log works: each night's labels are appended to () as one write-once file per date. A manifest records each file's SHA-256 and the previous line's hash, so an edit to any earlier day breaks the chain.

Research and education only, not investment advice. Data may be delayed or wrong. Past performance does not predict future results.

License, security and privacy

• License: the code is MIT-licensed (LICENSE). Market data, news and other third-party content keep their own terms; the website's privacy and data-sources page lists each source and credits it.
• Security: report problems privately, as described in SECURITY.md. No keys or secrets belong in the repository; local settings live in .env, which git ignores.
• Privacy: the website has no server, accounts, cookies or analytics. Settings stay in the visitor's browser, and the privacy page lets visitors clear them. The page also lists every service the browser contacts.
• Tests on every push: .github/workflows/tests.yml runs python -m pytest on Python 3.14 with Node.js 24, for every push to main and every pull request.
  • It installs requirements.txt without torch and transformers (the tests use the lexicon sentiment scorer, so CI never downloads FinBERT), plus requirements-quantum.txt.
  • The dashboard smoke tests skip there, because there is no local database.
  • The actions are pinned to commit SHAs, and the job can only read the repository.
• Dependabot: .github/dependabot.yml opens weekly pull requests for outdated Python packages (Qiskit grouped on its own) and actions, and the tests check each one. The JavaScript libraries loaded from jsDelivr (Plotly 2.35.2, Lightweight Charts 5.2.1) are pinned in the HTML and JS and updated by hand.