Nasdaq-100 Removals — the year after leaving the index

A small data study of stocks removed from the Nasdaq-100 (NDX): for every removal in the last ~20 years (where the company kept trading), it measures how the stock behaved over its first 252 trading days out of the index — trough depth, peak, one-year return, and excess return versus QQQ over the identical window — then renders the results as a single self-contained, no-build static page styled with the Comet design system (dark, warm-purple, finance-grade).

• index.html — the whole site in one file. The overview has an auto-generated executive abstract, a coverage/survivorship panel with bias-corrected scenarios, the "average path" chart, outcome distributions, timing scatters, a tenure/archetype view, by-year and by-sector cohorts, and a sortable/filterable removals table (search · hide-truncated · year range · a 5-way market-regime toggle). Clicking a removal opens a per-stock deep-dive overlay — verdict KPIs, indexed price path vs QQQ, a market-regime return decomposition, a quarter-by-quarter timeline, commentary, and a comparable-removals table.

The deep-dive overlay is synced to the URL hash, so links like index.html#NFLX-2012-12-24 open a specific stock directly and are shareable. Every chart supports wheel/pinch zoom, drag-to-pan, and double-click reset.

Repository layout

data/
  raw/         removals.csv            hand-reviewed Wikipedia scrape (the pipeline input)
  processed/   results_per_stock.csv   per-stock summary stats (study output)
               series/                 daily aligned stock/QQQ series cache (+ _qqq, _manifest)
               universe.csv            full 204-removal universe + fate (survivorship base)
               fates_to_review.csv     unresolved fates for hand-filling (re-read on rerun)
               sectors.csv             GICS sector per ticker (correctable)
               *.json                  embedded aggregates: survivorship, average_path, tenure,
                                        cis, roundtrip, regime, sectors, comparables, abstract
src/
  ndx_removals_study.py   build the removal list + run the per-stock analysis
  export_series.py        fetch/cache daily series and reconcile them against the CSV
  build_universe.py  survivorship.py  average_path.py  tenure.py  bootstrap.py
                          pass-1 analysis: coverage/survivorship, average path, tenure, bootstrap CIs
  roundtrip.py  regime.py  sectors.py  comparables.py  abstract.py
                          pass-2 analysis: fate/round-trips, market regime, sector, peers, abstract
  build_site.py           render data/processed -> dist/index.html (the whole Comet site, self-contained)
  comet-tokens.css        Comet design tokens (inlined into the page by build_site.py)
  qqq-favicon.svg         favicon (embedded as a data URI by build_site.py)
  wrapup.py  wrapup2.py   reconciliation + headline reports (read-only)
dist/          index.html  the built site (gitignored — regenerate from src/)
logs/          run logs (gitignored)
requirements.txt  README.md  LICENSE  .gitignore

The processed data (results_per_stock.csv, the series/ cache, and the embedded aggregates) is committed so the site can be rebuilt offline — delisted-ticker history is hard to refetch later, so the cache is the source of truth for the pages.

Setup

pip install -r requirements.txt      # tested on Python 3.13

Pipeline

Run the scripts from the repository root. They resolve their own paths relative to the repo, so the working directory doesn't matter, but the src/ prefix below assumes you're at the root.

1. Build the removal list (needs internet — scrapes Wikipedia):

   python src/ndx_removals_study.py build-list

   Writes data/raw/removals.csv. Review it by hand: verify dates and the include/reason columns, since acquisition- and delisting-driven removals (no real "year after") are excluded. Cross-check against Nasdaq's December reconstitution notices.

2. Run the analysis (needs internet — pulls daily data via yfinance):

   python src/ndx_removals_study.py analyze

   Reads data/raw/removals.csv, writes data/processed/results_per_stock.csv, prints the macro summary, and reports coverage (a low coverage rate means survivorship bias — see the caveat in the script docstring).

3. Cache + verify the daily series (network only for tickers not already cached):

   python src/export_series.py

   Writes data/processed/series/*.json and re-derives each stock's stats through the same code path that produced the CSV; it exits non-zero if any recomputed stat disagrees with the CSV beyond tolerance, so the pages never ship inconsistent numbers.

4. Compute the aggregates (offline — reads the processed data; later steps read earlier outputs):

   python src/build_universe.py     # universe.csv + fates_to_review.csv  (hand-fill unresolved fates, then re-run)
   python src/survivorship.py       # survivorship.json  (scenario medians + bias range)
   python src/average_path.py       # average_path.json  (offset-aligned median/IQR bands; reconciles vs the CSV)
   python src/tenure.py             # tenure.json  (years-in-index + archetype; scrapes Wikipedia)
   python src/bootstrap.py          # cis.json  (10k bootstrap 95% CIs)
   python src/roundtrip.py          # roundtrip.json  (re-additions + ultimate fate; scrapes Wikipedia)
   python src/regime.py             # regime.json  (market regime per window; reconciles QQQ vs the CSV)
   python src/sectors.py            # sectors.csv/json  (GICS sector; edit sectors.csv to correct)
   python src/comparables.py        # comparables.json  (similar prior removals per stock)
   python src/abstract.py           # abstract.json  (grounded executive summary)

5. Build the site (offline — reads only the processed data):

   python src/build_site.py

   Reads data/processed/* directly and writes the single self-contained dist/index.html (Comet tokens inlined, favicon embedded). Open it directly in a browser (Chart.js, the zoom plugin, Lucide, and Google Fonts load from CDNs — the only things that need internet when viewing).

   Optional: python src/wrapup.py / wrapup2.py print reconciliation spot-checks and the headline numbers.

Note: dist/ is gitignored (a deploy artifact), so it isn't in the repo — your locally built/served dist/index.html is the live site, and build_site.py reproduces it 1:1 from the processed data.

Deploying

The built site is a single self-contained file — copy it to your server and serve it:

index.html      (the whole site: overview + per-stock deep-dive overlay)

• There are no per-stock files and no sidecar assets — the design tokens, favicon, and all per-stock daily series are embedded in index.html. Each deep-dive is a URL-hash route within it (e.g. index.html#ZM-2023-12-18), so those links are shareable.
• Do not upload data/, src/, or logs/ — they are build-time inputs only and are never read at runtime.
• Viewing needs outbound internet for the Chart.js / zoom / Lucide / Google Fonts CDNs; nothing else is fetched.
• index.html is ~1 MB. Upload it in binary mode and confirm the server copy's byte size matches your local file. A truncated/corrupted transfer is the usual cause of a blank page (the embedded data breaks and the script aborts before rendering); check the browser console for a SyntaxError if it comes up empty.

Data & attribution

• Price data via Yahoo Finance (the yfinance library); a partial Stooq fallback is used only where Yahoo has no data, flagged in the source column.
• The removal list is derived from the Wikipedia "Nasdaq-100" article's component-change tables.
• "Nasdaq-100" is a trademark of Nasdaq, Inc.; "QQQ" (Invesco QQQ Trust) is a product of Invesco — referenced here for identification only. This project is not affiliated with, endorsed by, or sponsored by either.
• Survivorship caveat: tickers that were delisted or acquired after removal are dropped by the data source, and those were disproportionately the worst performers — so every figure is biased upward. The pages state this prominently.

Built by Hesanka with Claude.

License

Code is released under the MIT License. The license covers the code in this repository, not the third-party market data it fetches, which remains subject to its providers' terms.