Universe Builder

universe-builder

You hand over a filter chain. The skill walks a candidate universe of US stocks, applies each filter in order, ranks the survivors by a composite z-score, flags sector concentration in the top decile, and emits two output layers from one analysis.

This is the lowest-barrier skill in the suite. Free Basic keys can run it end to end (with throttling) on a small candidate set. Paid keys can run it across thousands of names without rate-limit pain. The output is the starting point for everything else: factor-research backtests, pitch-comps, options-flow watchlists.

Unlike a generic screener, universe-builder:

• Records every step of the filter chain so the survivor count is auditable, not "trust the dashboard"
• Computes a composite z-score across surviving factors, not just single-factor sorts
• Flags sector concentration in the top decile (the screen result you get often has more sector skew than you expected)
• Documents survivorship handling explicitly (delisted names retained for any historical lookback)

When to invoke

• A PM says "give me US large-caps with strong momentum and decent cash yield"
• A quant says "starting universe for a 3M momentum backtest, no micro-caps, no illiquid names"
• A discretionary analyst says "what are the top 20 names by 3M momentum that also have positive operating cash flow"
• A retail user says "screen US stocks above $10B, filter by mom and valuation"

What you need

• A filter chain (CLI flags or a JSON config)
• MASSIVE_API_KEY exported in the environment
• Stocks Basic plan minimum (free works end to end at small candidate sizes; paid removes the 5/min rate cap)

Canonical CLI flags

The screener-style flags all use the --<bound>-<factor> shape so a filter chain reads top to bottom like a SQL WHERE:

--min-price 20             # last close >= $20.00
--min-adv 400000           # 20d avg daily volume >= 400,000 shares
--min-mom-3m 0.10          # 3M momentum >= +10% (canonical)
--max-week-return 0.0      # 5d return below threshold (see semantic below)
--min-mcap 10e9            # market cap >= $10B
--max-mcap 100e9           # market cap <= $100B
--ocf-yield-min 0.03       # operating CF yield >= 3%
--include-sectors X,Y      # categorical sector include
--exclude-sectors X,Y      # categorical sector exclude
--include-types CS         # security type whitelist (default 'CS', see below)

--mom-3m-min is a deprecated alias for --min-mom-3m. It still works but prints a warning. Use the canonical --min-mom-3m so the flag style matches --min-price and --min-adv.

--max-week-return semantic

The threshold is signed and the comparison operator changes at zero so the natural-language reading matches the math:

• --max-week-return 0.0 keeps names where week return is strictly less than 0 (excludes flat names — we want pullbacks, not stasis)
• --max-week-return -0.05 keeps names where week return is at or below -5% (the user-intuitive reading of "down 5% or more" includes the exact -5.0% case)

In other words: zero is strict <, negative thresholds are inclusive <=.

--include-types (default CS)

The default keeps common stock only so a "stock screen" doesn't silently include ETFs, leveraged products, foreign ADRs, or warrants. The enrichment pass (see below) tags every survivor with its Massive type before this filter runs.

Override examples:

• --include-types CS,ADRC — also include foreign ADRs (LEGN, etc.)
• --include-types CS,ETF,ETN — include ETFs and exchange-traded notes
• --include-types '*' — disable the type filter entirely

Massive returns these types in practice: CS (common stock), ETF, ETN, ETV (ETF / ETN / ETV variants), ADRC (foreign ADR), PFD (preferred), WARRANT, RIGHT, UNIT, FUND. The list endpoint exposes the filter, but the type field is only populated on per-ticker details, which is why the skill defers the type filter to the enrichment pass.

Enrichment pass

After the cheap price / volume / momentum filters reduce the working set from ~12,000 names to 300-2,000, the skill makes a parallel fan-out of per-ticker /v3/reference/tickers/{T} calls (16 workers) to pull type, sic_code, sic_description, market_cap, and the human name. Without this pass the grouped-aggs path has no security-type data, so leveraged ETFs and 2x products leak through and the concentration check shows everything as "Unknown".

Cost on Business tier is under 30 seconds for a 345-name cohort. Massive's list endpoint silently ignores ticker.any_of=... for batch lookup (probed 2026-06-24), so per-ticker fetches in parallel are the right shape.

The skill runs at two fidelity tiers, flagged in the output JSON as tier.

• Tier A (paid Stocks Starter or above): Unlimited REST. The candidate pool can be a few thousand names; the full filter chain runs in under two minutes.
• Tier B (free Basic): 5 calls/min. The candidate pool defaults to 100 names (top US mcap by curated seed). Documented as the on-ramp, not the production configuration.

What you get back

The skill ships two output layers from one analysis.

Layer 1: canonical JSON matching output-schema.json. The filter chain steps with per-step survivor counts; the surviving rows with their per-factor z-scores and composite rank; the sector concentration analysis with std-devs overweight per dimension; the source endpoints with fetched-at timestamps. UIs, downstream agents, and Python scripts consume this.

Layer 2: rendered table in Bloomberg EQS / FactSet screener style. See references/rendering.md for the canonical format rules. Monospaced columns, signed percentages, a filter survival funnel as a separate small table, a concentration callout when the top decile is sector-skewed, and a survivorship line at the bottom.

UI devs build their own table from the JSON. Claude Code users see the rendered form.

How it works

1. Build the candidate pool. On Tier A, paginate /v3/reference/tickers?market=stocks&active=true until the cap is hit. On Tier B, use a curated seed list of large-cap tickers to stay under the rate limit.
2. Apply filters in cost order per references/filtering-methodology.md. Cheap predicates (active flag, exchange, type) come first; expensive ones (multi-day momentum, financials) come last. Each step records its survivor count so the funnel is auditable.
3. For each surviving name, fetch market cap and sector from the ticker details endpoint, momentum from a daily aggregate window, operating cash flow yield from the quarterly financials endpoint (FCF requires CapEx which Massive doesn't expose separately, see references/filtering-methodology.md), and options ADV from the contracts list if Options Developer is available.
4. Compute per-factor z-scores within the surviving universe per references/composite-zscore.md. Each factor is signed per direction (higher mcap = better, higher FCF yield = better, lower P/E = better), then equal-weighted into a composite z-score that drives the final rank.
5. Detect sector concentration in the top 20 names per references/concentration-analysis.md. If a sector is >2σ overweight vs its share of the starting universe, flag it.
6. Document survivorship per references/survivorship-handling.md. The active flag from Massive's reference endpoint indicates whether a ticker is currently trading; backtests should set active=false in the lookback window to avoid the standard bias.

Foundations used

• massive-api-patterns for REST auth, rate limiting, and the best-price fallback chain

Output mode: table

Table mode is the format Bloomberg EQS, FactSet screeners, and any quant blotter use to compare a ranked set of names on a fixed set of columns. Monospaced or markdown-table layout so values line up; the filter survival funnel renders as a second small table; the concentration callout renders as bullet lines.

references/rendering.md is the canonical format reference for any future table-mode skill (factor-research, pitch-comps). Match this format.

Endpoints used

• GET /v3/reference/tickers?market=stocks&active=true: paginated candidate pool. Cheap; one-call per page.
• GET /v3/reference/tickers/{ticker}: per-name market cap, sector (SIC code), industry, shares outstanding. One call per surviving name.
• GET /v2/aggs/grouped/locale/us/market/stocks/{date}: one call per date returns OHLCV for all ~12,000 active US stocks. Used to compute momentum efficiently (two calls: today and N days ago).
• GET /vX/reference/financials?ticker={ticker}&timeframe=quarterly: per-name TTM operating cash flow for the cash-flow yield factor. Free Basic includes this endpoint.
• GET /v3/reference/options/contracts?underlying_ticker={ticker}: per-name options ADV, when Options Developer is available. Optional factor; skipped on Stocks-only plans.

Doesn't handle (yet)

• True free cash flow yield. Massive's financials endpoint does not expose CapEx as a separate line, only net_cash_flow_from_investing which lumps CapEx with securities purchases and acquisitions. The skill uses operating cash flow yield as the most-defensible substitute and labels it as such. Add a separate CapEx parser keyed off source_filing_file_url (the raw XBRL) in a future PR.
• Forex / non-USD market caps. The skill restricts to US-listed common stock (type=CS, market=stocks).
• Custom factor weights. v1 uses equal-weighted composite z-score. Per-factor weights are wired into the schema (composite_weights) but not exposed as CLI flags yet.
• Intraday filters. Momentum windows are in trading-day units; intra- session screens belong to a separate live-screener skill.
• Cross-sectional industry adjustment. The concentration check is per-sector, not industry-adjusted within sector. A semi-heavy screen flags as semi-heavy, which is the right call most of the time.

These are clean PR extensions. The filter chain is composable by design.