Real-Time Machine Learning in the Cross-Section: A Public-Data Reproduction

Overview

Implementation

Core libraries

LightGBM, torch, scikit-learn, statsmodels, pandas, pyarrow, pydantic, Jinja2, openassetpricing

Replication scope

The target paper’s 207-anomaly Chen–Zimmermann sub-analysis over 1987–2019 out-of-sample, not the full signal universe

Protocol

Annual July-to-June prediction, recursive annual refits, fifteen boosted-tree fits per window before refit

Data sources

OSAP anomaly signals, yfinance monthly prices, Ken French factors, Hou–Xue–Zhang Q-factors, Chen–Velikov effective spreads, FRED macro, Wurgler sentiment

Configuration

33 YAML files: default, headline, smoke, public and local baselines plus 28 registered ablations

Tests

46 modules, including a fixture integration pipeline and a golden report test

Components

src/data/

One downloader per source — OSAP signals, prices, French and Q factors, Chen–Velikov spreads, Wurgler sentiment, Pástor–Stambaugh liquidity, FRED macro — each recording its vintage

src/data/reconcile.py

Security-identifier reconciliation between the price, signal and factor panels

src/data/panel.py, preprocess.py

Panel assembly and the cross-sectional rank transform to [−1, +1] with post-ranking imputation to zero

src/models/brt.py, linear.py, nn.py

The boosted-tree model and the linear and neural comparators behind one wrapper interface

src/analysis/protocol.py, train.py

The recursive annual training driver and the window protocol

src/analysis/portfolio.py, costs.py

Decile formation, long-short construction, and the effective-spread cost treatment

src/analysis/eval.py, subperiods.py

Factor-adjusted evaluation and the sub-period breakdowns

src/analysis/decompose.py, attribution_map.py

The gap decomposition attributing the replication difference to identified causes

src/analysis/ablation_defs.py, ablation_runner.py

The registered ablation set, its runner, and the selection cache

src/reporting/validation.py

The publication gate behind validate-production-run

src/utils/manifest.py, snapshot.py

The run manifest and input snapshot hashing

Stages

1. download-* Nine acquisition stages, one per source, each recording a vintage and writing a checkpoint.
2. reconcile-permno-ticker Resolve identifiers across the price, signal and factor panels.
3. build-universe, build-panel Assemble the point-in-time universe and the annual signal panel.
4. preprocess Cross-sectionally rank and impute signals under the declared schema.
5. train-brt Run the recursive annual fits; train-linear and train-nn supply the comparators.
6. form-portfolios Form deciles and broadcast annual predictions to monthly portfolio returns.
7. evaluate Compute factor-adjusted performance and the sub-period breakdowns.
8. run-ablations, decompose-gap Run the registered ablations and attribute the replication gap.
9. render-report, validate-production-run Render the note, then gate it.

Reproducibility and validation

• Every stage validates its output against src/schemas.py before writing, so a broken column contract fails at the stage that caused it rather than three stages later.
• Commands that are not yet implemented fail closed with a manifest entry marked not_implemented or blocked, instead of returning an empty result that looks like a finding.
• validate-production-run fails closed on smoke and fixture paths, missing artifacts, placeholder report text, bad snapshot hashes, and on the forbidden full-universe tag that would misstate the replication scope.
• The smoke configuration is a labelled 30-year, 100-security target with synthetic data and an injected regressor; it exercises the annual window protocol and is explicitly not a result.
• Known data limits are handled rather than ignored: the Chen–Velikov spread panel ends in 2017, so post-2017 cost reporting falls back to a declared fixed-basis-point sensitivity grid.

Availability