Public OSS mirror of the FX quantitative research framework. Bloomberg/enterprise adapters and secrets live only in the private repository Kozphy/ForexTraderWithBloombergAccess.

Reproducible quantitative research from raw market data to statistically validated, policy-controlled deployment.

This repository is a research and simulation framework with automated integrity tests. It does not claim live profitability, and CI runs on synthetic FX-like data so the pipeline is executable without proprietary Bloomberg credentials.

Raw Market Data
       ↓
Point-in-Time Data Layer
       ↓
Golden Schema / Data Quality
       ↓
Feature & Signal Registry
       ↓
Statistical Validation
       ↓
Walk-Forward / Time-Series CV
       ↓
Backtesting + Transaction Costs
       ↓
Portfolio & Risk
       ↓
Out-of-Sample Evaluation
       ↓
Policy Gate (PASS/FAIL evidence)
       ↓
Shadow / Paper Deployment
       ↓
Monitoring → Drift → Kill Switch → Audit

• Point-in-time dataset construction, schema validation, lineage and fingerprinting
• Causal (lagged) features and leakage rejection tests
• Cost-aware backtests with spread, slippage, commission, execution delay and liquidity caps
• Train / validation / OOS separation with purged walk-forward folds
• Configurable promotion policy that never silently promotes failures
• Control-plane primitives: human approval, retry budget, kill switch, JSONL audit
• Evidence bundles under artifacts/<experiment_id>/ with reviewer reports

python -m pip install -e ".[dev]"
pytest -q
python -m forex_quant.pipeline --config configs/baseline.yaml --artifacts artifacts

python -m forex_quant.pipeline --config configs/pass_demo.yaml --artifacts artifacts
python -m forex_quant.pipeline --config configs/fail_demo.yaml --artifacts artifacts

Inspect artifacts/exp_*/report.md and policy_decision.json.

On every push to private main, .github/workflows/sync-public-mirror.yml runs scripts/sync_public_mirror.sh and force-pushes a clean allowlisted tree to the public mirror. Secrets (.env, .secrets, tokens) and src/internal/ never leave the private repo.

Manual dry-run:

DRY_RUN=1 ./scripts/sync_public_mirror.sh

See architecture.md for the module map and control-plane lifecycle.

Supporting docs:

• research_methodology.md
• reproducibility.md
• benchmark.md — benchmark and failure-mode tables
• threat_model.md — threats to validity
• reports/RESEARCH_TEMPLATE.md

src/forex_quant/      # public research framework (mirrored)
src/internal/         # private-only Bloomberg/enterprise adapters (never mirrored)
configs/
  baseline.yaml
  pass_demo.yaml
tests/                # unit, data-contract, leakage, policy, audit
artifacts/            # generated evidence (local / CI; not mirrored)
evidence/             # curated pointers / gap analysis
scripts/
  sync_public_mirror.sh

GitHub Actions runs:

1. Unit tests
2. Data-contract tests
3. Leakage tests
4. Research harness / policy / audit tests
5. Smoke pipeline + artifact schema checks

A PR fails when research integrity requirements fail.

• Demo market data is synthetic; Bloomberg adapters are out of band.
• Statistical utilities are research building blocks, not independent model validation.
• Market-impact and capacity models are approximations.
• No unattended live execution path is provided.

Summarized in threat_model.md: leakage, survivorship, multiple testing, cost understatement, non-stationarity and irreproducibility.