Cross-Sectional Intraday Reversal at Multiple Horizons

Overview

Implementation

Core libraries

pandas, pyarrow, NumPy, SciPy, PyYAML, Jinja2, Matplotlib

Vendors

Polygon and Alpaca behind one vendor interface, plus a synthetic fixture vendor for offline runs

Analysis grid

5, 30 and 60-minute horizons, decile sorts with configurable sort gaps, and configured open and close minutes excluded from the headline

Configuration

A default YAML plus fixture, smoke and local configurations, with dotted --set overrides

Tests

19 modules, including a fixture integration pipeline and a golden report-HTML test

Scale note

A full production download runs to tens or hundreds of gigabytes depending on vendor coverage and retry history

Components

src/universe/

Point-in-time membership, the naive current-list comparator, and symbol and corporate-action overrides

src/data/vendors/

Polygon, Alpaca and fixture vendors behind one interface

src/data/ingest.py

Checkpointed, rate-limited, resumable bar ingestion with sidecar provenance and provider query logs

src/data/clean_bars.py

Streaming cleaning and the session calendar rules

src/analysis/returns.py, sorts.py

Horizon returns and the decile sorts

src/analysis/bootstrap.py

Block-bootstrap confidence intervals that respect intraday dependence

src/analysis/decay.py

Decay half-life fitting across horizons

src/analysis/costs.py

Turnover, the cost grid, and break-even cost levels

src/analysis/neutralization.py

Microstructure and sector controls

src/data/events.py

The optional earnings calendar behind the included/excluded sensitivity

src/validation.py

The publication gate behind validate-production-run

src/utils/manifest.py, repro.py

The run manifest and the determinism helpers

Stages

1. build-universe Construct point-in-time membership and resolve symbol overrides.
2. ingest-bars Download minute bars, resumably and under a rate limit; --dry-run plans without fetching.
3. clean-bars Clean and calendar-align, streaming rather than loading the whole panel.
4. compute-returns Build horizon returns with the configured sort gaps and edge-minute exclusions.
5. run-sorts Form deciles and long-short portfolios at each horizon.
6. bootstrap Block-bootstrap the headline spreads.
7. fit-decay Fit the decay half-life across horizons.
8. cost-overlay Apply the cost grid and compute break-even levels.
9. render-report Render the brief, figures and artifact tables.
10. validate-production-run Gate publication; exit non-zero on anything demo-labelled or incomplete.

Reproducibility and validation

• validate-production-run checks required artifacts, report sections, placeholder-free headline text, point-in-time membership presence, demo-mode guards, and demo or smoke labels in the report header — and is expected to exit non-zero on fixture output.
• The fixture path is explicitly synthetic and marked as such; demo results cannot be promoted to a headline because the gate refuses them.
• A one-symbol-week credential smoke run is the documented step before any full download, so a vendor problem surfaces in minutes rather than hours.
• A golden test pins the rendered report HTML, so a change in figures or tables has to be acknowledged rather than absorbed.
• The limits are recorded in the repository, not only in the write-up: the spread proxy is bar-derived rather than quote-derived, and the cost overlay is a round-trip turnover assumption, not an execution simulation.

Availability