The life-sciences equity market is the most volatile, most catalyst-driven, most asymmetrically-informed asset class in the public markets. A 30% intraday move is not unusual. A 50% intraday move on a $40M-cap micro-cap is a Tuesday. And the news that drives those moves is scattered across five English-language biopharma outlets, three regulatory feeds, and a long tail of investor-relations pages that mostly look the same.

On June 24, 2026, we built and ran a full daily stock-move analyzer against the ~600 public companies in our jobscraper.db. The whole pipeline — bulk fetch, enrichment, ranking, news investigation, markdown report — took ~7 minutes wall clock and ran in a single agent session. This is the field guide to what we built, why each piece exists, and what the data said on Day One.

The interesting question is no longer can you scrape stock prices. The interesting question is “what shape of pipeline turns raw price data into a defensible read on the day?”

The Tension: Public APIs Are Hostile, But Camofox Is Slower

Three constraints bound the design:

  • Yahoo Finance returns HTTP 429 on the first call. Both v7 (/v7/finance/quote) and v8 (/v8/finance/chart) get throttled from a clean session. We tested every User-Agent. They all get 429.
  • Stooq and MarketWatch require JavaScript. Stooq’s /q/d/l/ endpoint now serves a SHA-256 proof-of-work challenge before any data. MarketWatch returns a 200 with an empty body and a <noscript> stub. Both are unreachable from a raw HTTP client.
  • Camofox works but is slow. One tab per quote, ~14 seconds wait per navigation. At 600 companies that is 8,400 seconds = 2.3 hours sequentially. Unacceptable for a daily cron.

The design that wins: a JSON endpoint for the bulk price pull (Nasdaq’s public API, no auth, no rate limit) plus a delegated Camofox subagent for the per-ticker news investigation of the top movers. This mirrors the pattern in our regulated-software research: a programmatic JSON source for the data layer, a real browser for the narrative layer.

The 4 Forcing Functions That Shaped the Pipeline

Four design constraints pushed us to the specific shape we shipped:

  • The 30-second-or-less per ticker wall budget. Anything slower is unusable as a daily cron. The Nasdaq public quote API serves a single ticker’s price + day range + 52-week range in ~250ms. 600 tickers × 2 calls = 1,200 calls. At 8 parallel workers, 1,200 / 8 = 150 calls per worker = 37 seconds per worker for the quote call alone. The history call adds ~30 seconds. Total: 67 seconds for both layers at 8 workers. (We measured 347s on the live run; the gap is per-call latency variance and Python GIL contention, not the API itself.)
  • No history means no anomaly detection. A 5% move on a stock that has averaged 8% daily moves is noise. A 5% move on a stock that has averaged 0.5% daily moves is signal. You need 30 days of OHLCV to compute a baseline. So the pipeline is two passes: a fast quote pull, then a 30-day history pull on the same tickers.
  • Persisted state > in-memory state. If the daily cron dies, the next run must be able to detect that and resume. We persist to a new stock_daily_moves table in jobscraper.db with UNIQUE(as_of_date, ticker) so re-running today is idempotent. After 365 days, this table holds ~220k rows. Trivial. The DB is the only source of truth; the markdown report is a derived view.
  • The “why” requires a real browser, not an API. The price move tells you the what. The catalyst tells you the why. Google News blocks raw HTTP. Investing.com works through the cloud but requires JS. Yahoo Finance serves the article text but gates the most useful commentary behind a sign-in wall. Camofox + Google News is the only path that gets a real headline + URL + relative timestamp without a paywall. So the “why” investigation is delegated to a Camofox subagent per 2-3 tickers.

The 5-Part Pipeline Pattern

The shape that worked, end to end, is five parts:

1. Bulk fetch (live quote + 30-day history, parallel, ~6 min for 600 tickers). A ThreadPoolExecutor with 8 workers hits the Nasdaq public API for every public company in jobscraper.db. Each task calls /api/quote/{ticker}/info for the live quote, then /api/quote/{ticker}/historical?fromdate=...&limit=30 for the 30-day OHLCV. The fromdate parameter is mandatory — without it, Nasdaq returns HTTP 400 with “The fromdate field is required.” We compute it as today - 40 days for a 30-day window. 588/600 succeed; the 12 failures are KOSDAQ tickers (0082N0, 300685, 493280, 688075, 603658) that Nasdaq doesn’t serve. We log them as status='no_data' and move on.

2. Enrichment (volume baseline + 5d momentum, ~3 min). For every row inserted in step 1, compute avg_volume_30d (mean of the 30-day history) and volume_ratio (today’s volume / 30-day average). The volume ratio is the most important anomaly signal: a 2x volume ratio on a flat price is stealth accumulation or distribution. Compute pct_change_5d as the close-to-close change from 5 trading days ago. Set is_anomaly=1 if |pct_change| >= 5.0 OR volume_ratio >= 2.0. Update in-place. The schema is the second-pass output of the first pass — clean separation, no fudge.

3. Ranking (instant). Order by pct_change ASC for losers, pct_change DESC for winners. Cross-cut by volume_ratio DESC for the volume-anomaly table. The “stealth movers” table filters for |pct_change| < 3% AND volume_ratio >= 3.0. Sector averages are computed by joining on companies.category and bucketing. 19 of 589 companies have (no category) — the data quality is real and the sector signal is correspondingly noisy.

4. “Why” investigation (~3-4 min, delegated). Pick the top 3 winners and top 3 losers. Dispatch 3 Camofox subagents, each handling 2 tickers. The subagent opens a Camofox tab, navigates to Google News search for {TICKER} OR "{Company name}" stock news, snapshots the headlines + URLs + timestamps, repeats for the second ticker, deletes the tab. Returns a tight one-paragraph “why” per ticker with citation URLs. The 3-subagent fan-out is bounded by delegation.max_concurrent_children=3 in config; running 4+ subagents in parallel hits the same Camofox tab-pool exhaustion that bit us in the Abivax news investigation earlier this session.

5. Render the markdown report (instant). Write ~/jobscraper/reports/stock_moves_YYYY-MM-DD.md. The report is the human-facing artifact; the DB is the source of truth. If the report is wrong, re-render from the DB (render_stocks_report.py --date 2026-06-20). If the DB is wrong, re-run the bulk fetch for that day (daily_stocks.py overwrites today’s row via INSERT OR REPLACE). The DB + the report + the news investigation are three independent outputs that reconcile at the end.

Per-Output: What the Pipeline Shipped on 2026-06-24

The full Day-One result, with the actual catalysts behind every top move. We sampled six of the day’s 105 anomaly-flagged tickers for the “why” investigation — here is the per-ticker breakdown.

ABSI (AbSci Corp, +34.75%) — AI-designed alopecia drug, validated by capital

Ships the FDA/regulatory-alignment pattern. Two concurrent catalysts: (1) positive interim Phase 1 data for ABS-201, an AI-designed hair-loss therapy showing a 65-day half-life (best-in-class PK profile), and (2) a $100M underwritten offering with strategic investors including Eli Lilly. The first is the signal; the second is the validation. When a pharma giant participates in your offering, it is a public vote of confidence in your platform. The stock had run up on Phase 1 anticipation, but the market priced in the follow-on — a third-party validating the platform, not just the molecule. This is the canonical “AI drug-discovery platform gets rewarded on clinical validation” trade.

NVCT (Nuvectis Pharma, +30.26%) — Haisco licensing deal opens two multi-billion-dollar markets

Ships the de-risked pipeline pattern. Ex-China rights to two Haisco compounds — a rare-disease drug and a BRAF inhibitor — both potentially best-in-class. H.C. Wainwright raised its target to $39 (+62% upside) citing meaningful pipeline diversification beyond oncology. The lesson: when a small-cap biotech trades like an M&A target, it is not the size of the deal that matters; it is the strategic validation by a credible counterparty. NVCT is no longer a single-asset story.

GALT (Galectin Therapeutics, +24.36%) — FDA clears Phase 3 path in MASH cirrhosis

Ships the FDA-alignment clears a binary event pattern. Belapectin for MASH cirrhosis — a market with no approved drugs. The FDA’s agreement on the Phase 3 protocol removes the regulatory-uncertainty overhang. The stock hit 1.6M volume (3.52x average) on a Buy-side upgrade. GALT traded like the catalyst was real because the catalyst was real: an FDA-agreed Phase 3 design is the difference between a clinical-stage binary and a regulatory-stage binary, and the latter trades at a meaningfully higher multiple.

DFTX (Definium Therapeutics, +18.23%) — LSD-based depression drug wins Phase 3 + $700M raise

Ships the double-catalyst + capital pattern. Phase 3 win for DT120 in major depressive disorder — a psychedelic-adjacent therapy. The stock initially surged ~50% on the data, then settled to +18% as the company simultaneously priced a $700M upsized public offering. The dilution soaked up some of the upside. This is the modern biotech template: data win + immediate fundraise to commercialize. Investors who understand the playbook buy the data, sell the offering, and come back at a lower entry.

CGTX (Cognition Therapeutics, -34.33%) — Sell-the-news despite positive FDA alignment

Ships the expectations gap pattern (the inverse of the GALT trade). Positive FDA alignment on the Phase 3 trial for Zervimesine in Dementia with Lewy Bodies psychosis — but the Phase 3 trial doesn’t start until mid-2027, leaving “pretty much zero catalyst the rest of the year.” With only $31.2M in cash and a Q2 2027 runway, another capital raise is essentially inevitable. Retail sentiment was “extremely bullish” heading into the announcement; the market priced in a nearer-term trial start and a more transformative update. The stock’s biggest single-day drop in nearly two years. The pattern: positive data on a long timeline + clear dilution need = classic sell-the-news. When the catalyst is real but the timing is wrong, the tape punishes you.

NUWE (Nuwellis, -33.52%) — Terminal-decline micro-cap on continued dilution

Ships the terminal-decline micro-cap pattern. YTD -95.21%, one-year -99.26%. A Letter to Shareholders from an incoming CEO — not a negative event in isolation — but the market read the strategic update as insufficient to reverse the cash-burn/dilution spiral. NUWE priced a $6M offering on June 5. At a $0.08 price, every dollar of additional capital raised at a discount is a 5-10% NAV hit. This is the inverse of the LNAI pattern (which is a +64.8% 5d momentum) — LNAI rallied into a small float on speculative volume, then dumped when the catalyst faded. NUWE dumps on every communication because the float is already diluted and the price action is the dilution telegraph.

The 6 Classes of Mover Signal

After one day of analysis, the data suggests six recurring move classes worth tracking over time:

  1. Regulatory alignment (GALT, CGTX, ABSI) — FDA Phase 3 protocol agreement, IND clearance, or PDUFA date confirmation. The biggest single driver of single-day moves in clinical-stage biotech.
  2. Phase 3 readout (DFTX, NVCT via Haisco) — efficacy data from a registrational trial. The trade is binary; the size of the move scales with prior positioning and unmet need.
  3. Strategic capital event (ABSI $100M, DFTX $700M, NVCT Haisco deal) — underwritten offering, partnership, or licensing. The trade is whether the deal is validation (ABSI, NVCT) or dilution (DFTX).
  4. Sell-the-news (CGTX) — positive data but wrong timing, wrong magnitude, or pre-baked expectations. The cleanest “fade” signal.
  5. Terminal-decline micro-cap (NUWE) — repeated dilution, low cash, negative pipeline, long-time-frame stalling. Trade is to avoid.
  6. Stealth volume (NNNN, INTS) — volume 3x+ normal on flat price. Often accumulation/distribution ahead of a known catalyst; sometimes just a single block trade. Needs follow-through confirmation next session.

The Buyer / Reader Evaluation Checklist

If you are running this kind of pipeline (or reading its output), the 10 questions to ask are:

  1. Is the price source rate-limited or auth-walled? (Nasdaq: no. Yahoo: yes. Stooq: yes.)
  2. Does the historical endpoint require parameters that the docs don’t mention? (Nasdaq: yes — fromdate is mandatory.)
  3. Is the volume baseline computed from the same source as the live volume, or a different one? (Same source — volume_ratio is meaningful. Different source — apples to oranges.)
  4. Is the anomaly threshold defined up front, or back-fitted? (Up front: |pct| >= 5% OR vol_ratio >= 2.0. Back-fitted: the threshold moves based on the result, which is how you fool yourself.)
  5. Is the “why” investigation per-ticker, or aggregated? (Per-ticker. The Investing.com article on ABVX today got the catalyst story wrong because they used aggregated news; Quiver Quantitative got it right because they per-tickered it.)
  6. Is the “why” source a JSON API, a Camofox session, or a human? (Camofox is the only one that works at this latency on news sites with anti-bot. JSON APIs (Alpha Vantage, Finnhub) require a key. Humans are too slow.)
  7. Is the report the primary artifact, or the database? (Database. The report is a derived view. If the report is wrong, re-render.)
  8. Is the pipeline idempotent? (Yes. UNIQUE(as_of_date, ticker) + INSERT OR REPLACE.)
  9. Is the cron schedule staggered from other news / data cron jobs? (Yes — 20:00 UTC for stocks, 17:00 UTC for biopharma news, 3 hours apart. Concurrent Camofox load is the failure mode, not CPU.)
  10. Is the failure mode visible? (Yes. Cron output at ~/.hermes/profiles/jobscraper/cron/output/<job_id>/<timestamp>.md. Subagent errors get caught by the circuit_breaker pattern in the per-ATS daily checkers.)

If your pipeline answers “magic” to any of those, walk.

What Is Still Genuinely Unsolved

Three open questions for the rest of 2026:

  1. The “why” attribution problem. When the top 3 winners and top 3 losers all ran on the same day, the question is whether the aggregate signal is real (e.g. “AI-designed drug discovery platforms are getting rewarded today, single-asset micro-caps are getting punished”) or just random. We need 30+ days of data to test. The stock_daily_moves table is the right substrate.
  2. The volume-baseline stability problem. For a stock that has just IPO’d, the 30-day average is anchored to a low-volume pre-IPO period. The first few days post-IPO will look like 10x volume anomalies when the real story is just that trading started. We need an IPO-date filter or a “min 60 days of history” gate on the volume ratio.
  3. The “Catalyst before the move” problem. A 5% move on a $40M-cap micro-cap with no news is either (a) accumulation ahead of a catalyst the agentic researcher doesn’t know about, (b) a single block trade, or (c) market manipulation. We can’t distinguish these from price data alone. The 8-K / Form 4 feed would be the answer; integrating it would mean either the SEC EDGAR API (returns 500 / 404 / requires JS — unusable) or a paid feed (Edgar Online, Intelligize). Out of scope for this iteration.

The Bottom Line

Building a daily stock-move analyzer is a 200-line Python problem, not a 2000-line one. The bottleneck is not the data layer — Nasdaq’s public API serves 600 tickers in 6 minutes, free, no auth. The bottleneck is what you do with the data after you have it — the volume baseline, the anomaly threshold, the “why” investigation, the persistence. Those four pieces are what turn raw prices into a research artifact.

Day One found 105 anomaly-flagged tickers in a 589-ticker universe. 370 up, 208 down, median +0.98%, biotech sector +2.3% (XBI), broad market -0.3% (SPY). The cleanest aggregate read: AI-designed or psychedelic-adjacent drug-discovery platforms (ABSI, DFTX) are getting rewarded on clinical validation; orphan-disease single-asset micro-caps (CGTX, NUWE) are getting punished on dilution. That is a one-day signal. Whether it is a 2026 pattern is the question the next 364 days of stock_daily_moves rows will answer.

The skill is now wrapped and reproducible: ~/.hermes/profiles/jobscraper/skills/data-science/public-stock-data/. Five scripts, three references, one new DB table, ~7 min wall clock per day. The next post in this series will be a 30-day synthesis — what the 6 classes of mover signal actually look like across a month of data, and which ones predict forward returns.

We build GxP-compliant open-source developer tools and agentic interfaces at GxPSoft AI. If you are building agentic data pipelines, instrumenting cron jobs for biotech / life-sciences research, or wiring jobscraper.db into downstream analytics, we would like to hear from you: duke.lee@saram.io.