Screeners

A screener is a filtering strategy that turns the broad market into a short list of candidate symbols based on numerical criteria — financial ratios, price patterns, insider activity. It’s the discovery layer of cents: where the candidate universe comes from before the orchestrator evaluates any of it.

This page covers the five starter strategies, how to use them, and the operating model. For the philosophical backdrop — why the discovery layer is a measurable, swappable component rather than a hard-coded worldview — see Operating principles.

Why a screener matters

Without a screener, your universe is whichever symbols you happen to have heard of — a tiny biased sample of the ~5,000 US-listed stocks. Big-cap retail-attention names (AAPL, NVDA, TSLA) dominate; everything else is invisible. A screener lets the system systematically discover names you’d never have thought to research, that match a hypothesis you can articulate.

The classic flow is:

1. Hypothesis — “oversold quality companies bounce”
2. Screen — RSI(14) < 30 AND positive TTM ROE
3. Output — 5 to 15 names you’ve never heard of
4. Research — what the orchestrator does for each candidate
5. Position — what the factory opens for the qualifying ones

Without step 2, you can’t handle the volume — so you ignore most of the market. With it, the system can comb the universe systematically.

The five starter strategies

Each strategy maps to a recognizable school of stock-picking with decades of academic literature behind it. Default thresholds are intentionally selective — you’ll see what each one returns and tune from there.

value

Rule: P/E < 15 AND debt_equity < 0.5 AND ROE > 10% AND latest quarterly revenue growth > 0

Classic Graham value with quality and growth guards. In a market where the S&P 500 trades around 20-25× earnings, P/E < 15 is intentionally strict — typically returns 30-50 names across the S&P 500, heavily tilted toward financials, energy, and some healthcare. The ROE and revenue-growth gates prevent “value trap” names.

growth

Rule: 3y revenue CAGR > 15% AND latest gross margin > 40% AND gross margin nondecreasing over 3y

Real growth bar: about 10% of the S&P 500 clears 15% CAGR consistently. The 40% gross margin tilts toward software, brands, and pharma. The “nondecreasing margin” gate is the smart part — it strips out cyclicals masquerading as growth.

momentum

Rule: latest close > 50d SMA AND 3m price change > 10% AND recent 5d volume avg > 1.5× trailing 50d avg

Trend-following with volume confirmation. 3m at +10% catches sustained movers but misses early breakouts. Volume gate separates real moves from drift.

mean_reversion

Rule: RSI(14) < 30 AND positive TTM ROE

Textbook oversold quality, not raw oversold. The RSI threshold is rare — typically returns 0-5 names on any given day across the S&P 500. The quality gate is essential: without it, this becomes “buy falling knives.”

insider_cluster

Rule: ≥3 distinct insiders bought in last 30 days AND net insider dollar flow > 0

Cluster buying with net positive insider conviction. The “net positive” condition (rather than “zero offsetting sells”) tolerates routine 10b5-1 scheduled sales while still requiring genuine buying weight.

Universe scope: --over is the right default

Two ways to point a screener at the market:

# Constrained — runs the screen against a parent universe
cents universe create cheap_sp500 --source screener \
  --strategy value --over sp500

# Full-market — opt-in, env-flag gated
CENTS_SCREENER_ALLOW_FULL_UNIVERSE=1 cents universe create deep_value \
  --source screener --strategy value

The constrained form is what you’ll use 90% of the time — it’s faster, cheaper, more interpretable, and lets you compare screeners on the same parent set. The full-market form is gated behind an env flag so a cron job can’t accidentally launch a 5,000-symbol scan.

Common parent-universe patterns:

# Value within the S&P 500
cents universe create value_lc --source screener --strategy value --over sp500

# Momentum within your watchlist
cents universe create momo_watch --source screener --strategy momentum --over my-watchlist

# Multiple thematic baskets, screened differently
cents universe create energy_value --source screener --strategy value --over energy_basket
cents universe create semis_momentum --source screener --strategy momentum --over semis_basket

CLI

cents screener list                          # show all registered strategies
cents screener describe value                # print the rule set
cents screener preview value --over sp500    # dry-run, show what it'd produce

cents universe create <name> --source screener --strategy <s> [--over <parent>] [--limit N]
cents universe refresh <name>                # re-run the screen, update symbols

How screeners close the loop

Every thesis the factory opens records its discovery_source — the universe name (and hence the screener) that produced its symbol. That’s what enables the second feedback loop:

cents factory analyze --by discovery              # win rate, avg P&L by screener
cents factory analyze --by discovery,cohort       # 2D cross-tab
cents factory analyze --by discovery,regime       # regime-conditional

With this in place, the winning algorithm isn’t “which stocks to buy” — it’s a map:

“In regime R, use screener S to populate a universe of N candidates, evaluate them with the orchestrator, open paired theses with premise tags T, and manage with target / stop / horizon parameters P.”

That map gets sharper as outcomes accumulate. The screener layer is what makes the discovery half of it learnable instead of fixed.

Tuning defaults

Thresholds live in source_config.params, configurable per-universe:

cents universe create value_loose --source screener --strategy value \
  --over sp500 --params 'pe_max=20,roe_min=0.08'

You can have value_strict (P/E < 15) and value_normal (P/E < 20) running in parallel and let outcome data tell you which produces theses that actually win.

What it deliberately doesn’t do (yet)

• No custom screener DSL — start with the 5 built-ins; see what’s missing; design a DSL once we know what’s actually needed.
• No auto-refresh schedule — cents universe refresh is manual or cron.
• No multi-screener composition — no intersect(value, momentum) yet; separate universes do the same job at v1 scale.
• No backtest mode — point-in-time fundamentals is gnarly; current screeners run against current data only.
• No performance benchmarking against indices — you can compare screener-cohort outcomes via factory analyze, which is the meaningful comparison anyway.