Two New Features, Zero New Data Feeds

The Asset Both Features Stand On

Every night the public ticker page builder grades ~2,500 companies and writes each one's full summary (grade, score, sector, analyst sentiment, ML health, fundamentals, multi-year grade trend) as JSON into publictickers. That's the global grade cache the free tier already reads from. Next to it, the EOD price cache (fmppricecache) holds years of daily closes for every tracked symbol, kept warm by the cache warmer and every report run. The two features below are mostly new reads of those two tables.

Watchlist Dashboard: Where Each Column Comes From

One sortable grid of every symbol on the account's lists: price, day change, volume, 52-week range and position within it, RSI(14), analyst sentiment, health score, grade and a 30-day sparkline, with auto-refresh off / 1 / 5 / 15 minutes.

• Fan-in, not per-user polling. A process-wide quote memo (WatchlistQuoteCache) sits in front of the proxy with a 60-second TTL, equal to the fastest refresh option, so a 1-minute viewer never sees a quote older than the interval they chose. A hundred users watching AAPL cost one quote call a minute, not a hundred. Misses (delisted, OTC, typos) are cached too, so a bad symbol doesn't trigger a provider call on every refresh.
• The RSI matches other charting tools. Wilder's RSI, seeded with a simple mean and then Wilder-smoothed, which is the definition charting platforms display. It's pinned against a reference series in the tests. With fewer than 15 closes it returns nothing, because a short-window RSI is a different, noisier number. Today's quote is appended only if it's for a day the EOD cache hasn't recorded, so the RSI and sparkline end at the price shown in the row.
• Missing sorts last in both directions. A missing RSI isn't "the lowest RSI". Ties break on symbol so an auto-refresh never reshuffles equal rows. That logic lives once, in a shared WatchlistGrid class used by all three apps.
• Bounded. A 250-symbol ceiling per request, so a large staff test list can't turn into a thousand-symbol quote sweep. If quotes fail, rows fall back to the last EOD close and are labelled as such.

Stock Screener: A LINQ Pass, Not a Search Cluster

Screens run over the graded universe: grade score, sector, sentiment, P/E, dividend yield, revenue growth, debt/equity, ROE, FCF yield, operating margin, grade trend and ML health where it exists. The standard way to build this is a search index or a table of indexed filter columns. At ~2,500 rows, neither is worth it:

• A 10-minute in-process snapshot. ScreenerService parses the universe's JSON once and keeps the parsed rows in memory, replaced atomically so readers never see a half-loaded list. Grades change at most once a day, so a screen that's 10 minutes stale looks the same as a fresh one, and a search costs one in-memory pass instead of a ~10 MB database read.
• One definition of "yield ≥ 3%". Filtering, sorting, paging, sentiment grouping, the summary line and the criteria chips live in one shared ScreenerEngine (the same pattern as the Earnings Lab math), so the server and all three apps apply identical logic.
• Presets say their numbers. The built-ins (Undervalued Large Caps, High Growth Tech, Dividend Champions, Low Debt Leaders, Improving Grades) describe themselves in figures, e.g. "profitable companies above $2B yielding 3% or more with debt/equity of 2 or below". They're filters, not recommendations.

Hitting either free-tier limit logs a wall event (ScreenerResultCapHit, ScreenerPresetCapHit) into the same conversion telemetry as the other free-tier limits. Search works without the new screenerpresets table, so the feature degrades to "no saved presets" rather than failing if the migration lags the deploy.

What These Two Have in Common

Caveat: the screener's universe is the graded public set (the enabled index groups behind the public grade pages), not every listed stock. Widening it means grading more companies nightly, which is a cost and provider-contract decision, not a screener change.