NexusTrade Python SDK

Author trading strategies in typed Python. Backtest them on the engine that runs them live.

Quickstart · Authoring · Polling · Agents · Screener · Auth · Errors

---

pip install nexustrade

The base install is stdlib-only — no third-party dependencies, importable anywhere.

pip install 'nexustrade[stats]'   # spec curves, Newey-West, bootstrap
pip install 'nexustrade[documents]' # PDF inspection, OCR, structured extraction
pip install 'nexustrade[compute]' # exact union used by NexusTrade compute

Compute helpers such as nexustrade.host, nexustrade.scanned_table, nexustrade.signal, and nexustrade.report ship in this package. Host-backed operations require the short-lived environment supplied by a NexusTrade compute run.

Use nt.prepare_web_pages(pages, max_chars_per_document=1_000_000) for deterministic HTML inspection without a model call. pages maps source IDs to HTML strings/bytes, {"html": ..., "url": ...} objects, or successful host.fetch result rows. Each result has document and error; successful documents contain complete visible_text plus separate title, description and publication hints. The helper removes hidden content and page chrome and decodes HTML entities using the same parser as extract_web_pages and host evidence verification. Select contiguous report quotes from document["visible_text"]; do not concatenate metadata or separate passages. Over-budget text returns an explicit error, so a partial source is never presented as complete.

For source tables, nt.read_source_tables(original_html, source_id=...) preserves cells, spans, equal adjacent values and inline fact attributes. Inspect its grid; nt.select_source_table_cell(document, table=..., row=..., column=..., row_headers=[{row, column, text}], column_headers=[{row, column, text}], role=..., rationale=..., fact_attributes=...) asserts the exact source context. It raises on a failed selector rather than silently omitting a historical period. It does not choose accounting roles or establish absence. Prefer structured SEC role candidates/resolved facts when available; preserve selected records intact.

Both web helpers also accept successful fetch receipts typed text/markdown, including rendered fallback responses. These preserve the returned Markdown verbatim in visible_text and retain content_type; publisher metadata remains unset. Rendered text is evidence from that fetch response, not original HTML.

extract_web_pages supplies the complete prepared visible body, including notes outside main/article, within explicit character limits (default 200,000 per document and request). It never splices a page's beginning and end. Oversized sources return an error before a model call: inspect the retained source with prepare_web_pages, then explicitly raise affordable limits or select exact contiguous passages for narrower extraction. Successful results include host-owned coverage alongside document/error. This describes supplied text, not proof that the model correctly extracted every fact. Scripts, hidden text and page chrome are excluded by the visible-text projection; original fetched bytes remain the authority for inspecting those elements. A size error or a partial passage cannot support a claim that a disclosure is absent from the whole source.

For document-derived computation, keep extraction and interpretation separate. extract_rows/extract_pdfs preserve source observations. A corpus can recover document-level facts and logical rows in one schema-bound pass:

extracted = nt.extract_pdfs(
    documents,  # PDF bytes or successful host.fetch result objects
    document_schema={"report_date": "string", "filing_type": "string"},
    rows_schema={"asset": "string", "transaction_date": "string"},
    instructions="Return the rows requested by the task, preserving repeated rows.",
)

for source_id, result in extracted.items():
    document = result["document"]  # includes source_id
    rows = result["rows"]          # each includes source_id + _source_row_index

This is the beginning of the README. Read the rest at the original. Read the full README ↗