Building a stock screener inside Nova

I've spent enough years staring at TradingView screeners to know what bugs me about them. They surface lots of names. They tell you almost nothing about why a name passed the filter. They certainly don't tell you whether names like this one have made money historically. And once you find a candidate, you bounce to a second tool to write the trade plan.

The stock screener was supposed to be a small Nova feature. It became one of the most ambitious things I've shipped this year — and the one I use every day.

The pitch

A screener that:

• Ranks ideas by confluence across multiple independent layers — technical, fundamental, news/catalyst, options flow, social sentiment, macro regime.
• Is opinionated about direction: long or short, with the layers flipping accordingly.
• Generates a full trade plan per idea: entry, stop, target, risk-reward, primary catalyst, invalidation conditions.
• Backs every idea with historical backtest stats: setup-bucket win rate, avg R, expectancy, on a free-data daily-bar replay over five years.
• Costs nothing to run on day one. No paid data feeds. Free fallbacks for everything.

That last constraint is what shaped most of the architecture.

Free-data-first

Paid market data is great. It's also a hard pre-condition that excludes 90% of users (and me, sometimes). So the first cut had to work with free sources only:

• OHLC daily bars: Yahoo Finance's query1.finance.yahoo.com/v8/finance/chart. (Stooq's history endpoint went paid mid-implementation, which I learned on the day I tried to use it. Yahoo took its place; the portfolio code still uses Stooq's quote endpoint, which remains free.)
• Options chain: Yahoo, same domain.
• News: Google News RSS aggregating across publishers, plus Yahoo, Benzinga, Seeking Alpha RSS feeds.
• Filings: SEC EDGAR.
• Insider transactions: Finnhub when FINNHUB_API_KEY is set; soft-fail to neutral score otherwise.
• Social sentiment: StockTwits public message stream.
• Macro regime: FRED.
• Earnings calendar: Yahoo Finance.

Every layer is wrapped in a small Fetcher<TKey, TValue> contract that keeps the public-API shape stable. The day I pay for a Polygon subscription, the only thing that changes is which fetcher I plug in. The rest of the stack — confluence math, levels, backtest, UI — doesn't move.

Confluence and the directional flip

Each layer scores from 0–100 in long mode. Confluence is a weighted average. An idea passes the bar at 65+, which is conservative enough that a typical S&P 500 sweep returns 8–25 names, not 200.

The interesting wrinkle: long-bias and short-bias share the same scaffolding. The technical layer detects directional setups (mean reversion long, mean reversion short, momentum continuation, breakout, breakdown — five buckets). The screener picks the primary direction per ticker, then evaluates each layer in that direction's polarity. A bearish news layer is a positive contribution to a short setup and a negative contribution to a long setup. Same code, different sign.

Levels: ATR-based, R:R-gated

For every ranked idea, the screener computes:

• Entry: last close (the "ideal" entry; UI shows current price too).
• Stop: 1.5 × ATR from entry, clamped to the nearest support/resistance level if one exists within range.
• Target: 2× the entry-to-stop distance (the 2R default), unless an obvious S/R level caps it sooner.
• Risk-reward: the ratio.

Important behavior change shipped with phase 2a: cards with R:R below 1.5 still surface, but they don't get a trade plan. The idea might still be informative; the trade just isn't worth it. This was a deliberate choice — I'd rather see the underwhelming setup than have the tool silently swallow it.

Hybrid thesis: template by default, AI on demand

This is the one design decision I'm most proud of.

An AI thesis on every card sounds great. It is also expensive (every name in a 25-result scan = 25 LLM calls) and unreliable (the model can fabricate a date or misread a chart and you have to babysit it). A pure-template thesis sounds boring. It is also free, fast, and grounded in real numbers.

So Nova does both:

• Every card gets a deterministic template thesis — built from the actual signal layers, levels, primary catalyst from news, invalidation from levels. No LLM. Renders instantly. Numbers are guaranteed correct.
• Every card has an "Expand with AI" button — on click, the active brain rewrites the narrative and risk notes in prose, citing the same underlying data. Per-card, on demand. I pay for the AI thesis only on the names I'm actually about to take action on.

The hybrid means a 30-result scan is a $0 cost most of the time. When I expand three of them with AI, that's three calls. Not 30. I'll never go back to expand-everything-by-default.

Backtest engine

"This setup looks good" is opinion. "This setup has 0.11R expectancy across 413 trades over the last 10 years" is decision-grade information. So I built a small backtest engine that replays the technical signal layer over Yahoo's daily bars.

The engine takes a setup bucket, walks each ticker's daily history, fires a synthetic trade every time the signal triggers fresh (only on threshold cross from below — re-entries are not counted), applies entry/stop/target via the same levels.ts code the live screener uses, and tracks exit reason. Slippage is 0.1% of price. Max-hold caps per bucket: 10 days for mean reversion, 40 for momentum, 40 for breakout, 30 for misc.

Smoke run on 10 large-caps × 10 years: 413 trades, +0.11R expectancy, average hold 10.4 days, timeout exits under 13% across buckets. That's not Renaissance Technologies. It's also not noise.

Caveats I'm explicit about: technical-only replay (other layers don't have history I can reconstruct cheaply); daily-bar slippage simplification; current S&P-500 survivorship bias because the universe list isn't time-aware. Every limitation is in EDITORIAL.md-style docs in the project. I want to know what the number means before I trust it.

The UI

Initial cards: ranked grid. Filter row across the top — direction, setup type, min R:R, min confluence, symbol search, has-trade-plan toggle, has-historical-edge toggle. Click a card and a side rail opens with full detail: layered scores with explanations, levels visualized, news links, options unusual-activity flags, a per-card backtest button (replay this exact setup over history with one click), expand-with-AI button.

The Watchlist tab and Dismissed tab in the side rail let me triage. Click a name to watchlist it, click a name to dismiss it. Both tabs show the same full card so I can revisit decisions.

Every Detail view embeds a TradingViewChart with default RSI/MACD/EMA studies. I tried to do per-instance EMA20/50/200 and learned the hard way that custom-length studies require the licensed TradingView Charting Library. Default-param studies it is.

Persistence

Every scan writes:

• {vault}/Screener/scans/<timestamp>.json — the full scan output.
• {vault}/Screener/last-scan.json — pointer to the most recent.
• {vault}/Screener/Latest.md — markdown summary, opens nicely in Obsidian.
• {vault}/Screener/state.json — watchlist + dismissed list.

So my scans are part of my Obsidian vault now. They show up in graph view. They link to news items. The data is on disk in plain text in a folder I picked. Same principle as everywhere else.

What it shipped as

One screener skill. Eleven tools. One IPC namespace. One Dashboard tab. pnpm -r typecheck clean across 13 workspaces. Full S&P 500 large-cap scan with all five layers in about ten seconds. Fits in the same chat agent that does my calendar and reads my notes.

And then a few days later I started thinking about whether the screener could run itself. Which is how AI Picks happened — the autonomous trade-finder, in the next post.