Signal Research Framework

What this is

This is a sanitized reconstructed portfolio version of an offline signal research workflow. It demonstrates candidate generation, point-in-time checks, portfolio simulation, risk gates, selection, walk-forward evaluation, and reproducible reporting on synthetic data. The included metrics are research demonstration metrics, not platform scores.

Why I built it

The interesting engineering problem is not producing one impressive backtest. It is building a workflow that can say why a candidate entered or left the research funnel, prevent accidental look-ahead, control concentration and turnover, and reproduce the same result later.

Research workflow

Synthetic panel → textbook candidate grid → schema/static/runtime checks → duplicate removal → next-day backtest → metrics and fitness → risk gates → return-correlation filter → family-capped portfolio → walk-forward report.

Architecture

The src/signal_research/ package is split into signal generation, validation, backtesting, scoring, selection, risk, and reporting. configs/demo.toml holds all experiment thresholds. examples/run_demo.py is the end-to-end entry point. The tests cover each major contract and determinism.

AI-human collaboration

The human defines hypotheses, metrics, risk thresholds, and economic-plausibility judgments. AI generates candidate transformations, automates experiments, summarizes results, and detects duplicates. The system tracks provenance, evaluates, filters, and ranks. AI is a research multiplier, not a black-box strategy generator.

Signal lifecycle

Each immutable SignalSpec names a public textbook family, its required panel columns, numeric parameters, horizon, and rationale. Functions return a Series aligned to the full (date, ticker) panel. The engine applies date t weights to date t+1 returns and subtracts linear costs from one-way turnover.

Validation & risk gates

AST checks reject imports, file/network access, unsafe evaluation, and undeclared fields. Runtime checks verify exact index alignment and truncation invariance at multiple cutoffs spread through the sample. These checks are designed to catch accidental look-ahead and unsafe patterns in candidate code, not to sandbox a deliberately adversarial author — the funnel assumes good-faith research code and makes mistakes loud, not malice impossible. Risk gates then enforce minimum active observations, Sharpe, turnover, drawdown, and maximum absolute weight. A greedy correlation filter removes redundant return streams before family caps and top-k selection.

Example experiment

The checked-in example is generated with 100 synthetic tickers over 750 business days and a five-basis-point linear cost. The exact candidate counts, selected signals, metrics, and results_hash are in outputs/example_results/; they are refreshed by the commands below and are not live performance claims. In the verified run, the funnel was 54 candidates → 51 valid → 50 deduped → 13 passed risk gates → 6 correlation-filter survivors → 3 selected, with the selected signals landing between 1.24299 and 1.43186 annualized Sharpe after costs. The resulting hash was 2fb543d1a0faf4be42829acb0c947f770412ffb31b8dd2db6e20d9faa4fe9a94.

Reproducibility

Environment setup

From a clean checkout:

python -m venv .venv
./.venv/bin/pip install -e ".[dev]"

./.venv/bin/pytest
./.venv/bin/python examples/run_demo.py --config configs/demo.toml --out outputs/example_results

The run is offline and deterministic for a fixed configuration. lineage.json contains configuration, synthetic-data, stage, selection, and results hashes.

What is intentionally not included

There are no proprietary data fields, real competition signals or expressions, credentials, personal identifiers, external-service calls, platform scores, or claims of live profitability. The repository is a public educational reconstruction using synthetic data and textbook signal families.

WorldQuant-IQC disclosure

This is an independent sanitized reconstruction inspired by my private research workflow for the WorldQuant BRAIN / IQC competition; no real competition signals, platform data, or platform scores are included; not affiliated with or endorsed by WorldQuant. The metrics here are not platform scores.

Tech stack

Python, numpy, pandas, matplotlib, setuptools, TOML, and pytest. No network service or database is required.

Lessons learned

The most valuable controls are boring and explicit: a next-day return convention, a truncation test, canonical duplicate signatures, deterministic tie-breaking, risk gates, and lineage hashes. They make the research process easier to review and harder to accidentally overstate.