-orange){kind=icon}
QuantForge is a modular quantitative research framework focused on macro-conditioned trading systems across equities, FX, metals, and crypto. It features a live 24/7 paper-trading engine with a web dashboard and a complete model governance stack spanning sandbox evaluation, regime adversarial testing, trajectory tracking, and a formal promotion gate — turning model selection from metric comparison into regime-conditioned behavioral manifold analysis.
This is a research and paper-trading system — not a production trading bot.
The paper-trading engine runs continuously with the following allocation across 6 assets and 5 distinct driver clusters:
| Asset | Ticker | Label | Cluster | Alloc | Key Features |
|---|---|---|---|---|---|
| BTC | BTC-USD |
tb20 | momentum_crypto | 20% | rate_diff, 2y_yield_delta_63, vix_delta_5, dxy_mom_21, vix_ma21, btc_mom_{10/21/63}, btc_vs_spy_{21/63} |
| GC=F | GC=F |
fwd60 | real_asset | 20% | real_yield_delta_63, breakeven_delta_63, dxy_mom_63, gc_mom_63 |
| EURAUD | EURAUD=X |
tb20 | eur_cross | 22% | rate_diff, dxy_mom_21, vix_ma21, vix_delta_5, euraud_mom_{21/63} |
| NZDJPY | NZDJPY=X |
tb20 | carry_fx | 15% | vix_ma21, vix_delta_5, us_jp_10y_spread, nzdjpy_mom_21 |
| CADJPY | CADJPY=X |
tb20 | oil_carry | 13% | vix_ma21, ca_jp_spread_mom_21, us_jp_10y_spread, vix_delta_5, ca_jp_10y_spread, dxy_mom_21, cadjpy_mom_{10/21/63} |
| USDCAD | USDCAD=X |
tb20 | usd_macro | 10% | rate_diff, dxy_mom_21, vix_ma21, vix_delta_5, usdcad_mom_{21/63} |
./monitor_allDashboard: http://127.0.0.1:5000
- Engine refresh: every 30 minutes
- UI refresh: every 30 seconds
- Yahoo Finance resilience: exponential backoff (5s/15s/45s) + parquet cache fallback
- Portfolio summary (Total Value, Return, Unrealized P&L, Trade Count)
- Per-asset signal cards with confidence meters, position details, and P&L
- Session-scoped execution log (auto-cleaned on restart)
- Performance metrics (Profit Factor, Win Rate, Sharpe, Mean Confidence)
- Validity & Halt condition monitors
- Regime status and advisory bar
- Risk governance dashboard (
/risk.json,/risk/{asset}.json) - Shadow action intelligence (
/shadow-actions,/shadow-actions/{asset}.json)
Primary validated assets across 6 asset classes using walk-forward validation. Two label architectures are deployed: tb20 (triple-barrier, 20-bar timeout) for mean-reverting/short-term assets, and fwd60 (60-day forward return) for macro-trend assets. See ADR-015.
- Type: XGBoost multiclass classifier (BUY / NEUTRAL / SELL)
- Parameters: 300 trees, max depth 2, learning rate 0.02
- Labeling: Triple-barrier (pt_sl=2.0, vertical barrier=20 bars) for tb20 assets; 60-day forward return classification for fwd60 assets
- Sizing: Volatility-scaled positions
- Validation: Rolling walk-forward (5yr train / 1yr test / 1yr step)
5-year walk-forward evaluation: for each year (2021–2025), train a fresh model on data strictly before Jan 1, test on the full calendar year. Mean annualized Sharpe across the 5 out-of-sample windows:
| Asset | Label | Features | Avg Sharpe | Pos Windows | Avg DD | Avg Acc |
|---|---|---|---|---|---|---|
| USDCAD | tb20 | 6 | 1.48 | 4/5 | 0.016 | 0.5238 |
| NZDJPY | tb20 | 4 | 2.72 | 5/5 | 0.021 | 0.4717 |
| GC=F | fwd60 | 4 | 1.06 | 4/5 | 0.024 | 0.4049 |
| EURAUD | tb20 | 4→6 | 0.57→2.28 | 4/5→5/5 | 0.020→0.012 | 0.4816→0.4898 |
| BTC | tb20 | 4→10 | 0.39→0.83 | 3/5 | 0.286→0.144 | 0.4550→0.4182 |
| CADJPY | fwd60→tb20 | 4→9 | 0.08→1.70 | 3/5→4/5 | 0.026→0.011 | 0.3468→0.4876 |
Four assets (EURAUD, BTC, CADJPY, USDCAD) were augmented with additional macro and momentum features in May 2026. Arrowed values show baseline → augmented improvements. GC and NZDJPY were left unchanged (feature sets already optimal for their data-generating processes).
See ADR-015 for label methodology and ADR-016 for GC=F validation. See ADR-017 for inference integrity.
| Module | Description |
|---|---|
| HybridRegimeEnsemble | Global XGBoost backbone (depth=2, 100 trees) with regime-specific expert heads (TREND/RANGE/VOLATILE) and protected macro head (0.45 blend) |
| MacroExpertHead | Asset-specific XGBoost (depth=2) with protected weight 0.45 — prevents price features drowning macro signal |
| RegimeClassifier | TREND/RANGE/VOLATILE/NEUTRAL — operates as risk/participation controller |
| DriverAtlas | Asset-to-feature-cluster routing: carry_fx, yield_equity, momentum_crypto, positioning |
| ValidityStateMachine | GREEN/YELLOW/RED capital allocation with PSI-gated hysteresis |
| WalkForwardEngine | 5yr rolling train, 1yr OOS test, bootstrap p<0.05 deployment gate |
The feature algebra is governed by a 3-template grammar that covers 100% of deployed features. This eliminates train/serve skew by enforcing identical computation in both training and live inference:
| Component | Description |
|---|---|
| FeatureContract | Frozen dataclass — ticker, name, label type/params, macro filters, price momentum windows, vs-SPY windows. Runtime validators (validate_dataframe, validate_model) act as CI safety net |
| FEATURE_REGISTRY | Canonical dict mapping 6 tickers to their contracts — single source of truth for feature sets, label strategy, and model paths |
| Builder | build_features() — single generic pipeline: compute labels, align macro, compute template features, drop NaN. Called identically by training script and live engine |
| contract.features | Computed property — derives canonical column names from contract fields, eliminating hardcoded *_FEATURES constants and slug-based naming mismatches |
| model_path() | Returns {name}_model.pkl — fixes the old mismatch where training saved {slug}_model.pkl but engine expected {name}_model.pkl |
| Component | Description |
|---|---|
| FeatureContract / Registry / Builder | Canonical feature algebra — contract-driven feature computation, validation, and model path resolution |
| StateStore | Versioned persistence boundary — EngineSnapshot, corrupt-file recovery, file locking, cache management |
| StrategyRegistry | Per-asset strategy routing — model, signal, sizing, PnL, and feature pipeline interfaces with auto-register and validation |
| Shared Interfaces | shared/model.py, shared/signal.py, shared/sizing.py, shared/pnl.py, shared/features.py — ABCs + default implementations wrapping all critical computations |
| TradeDecision | Pure model intent dataclass — signal, confidence, provenance, no execution side-effects |
| PositionIntent / PositionManager | Extracted position lifecycle — open, close, SL/TP check, PnL accounting, deterministic replay |
| ValidityStateMachine | Per-asset GREEN/YELLOW/RED state with hysteresis, inertia smoothing, regime persistence lock |
| ExecutionState | Portfolio-level ACTIVE/PAUSED/HALTED — derived from halt conditions and validity exposure |
| AssetEngine | Per-asset XGBoost (depth=2, 300 trees, lr=0.02), contract-driven feature computation, tb20 or fwd60 label routing, strategy-registry-backed |
| PaperTradingEngine | Orchestrates 6 assets, volatility-scaled sizing, halt conditions, P&L tracking |
| PaperBroker(BrokerInterface) | Simulated broker — yfinance fills at market price, configurable slippage/fees |
| Dashboard (stdlib http.server) | Real-time web UI with portfolio summary, signal cards, session-scoped log, performance metrics |
| Wrappers | paper_trading/wrappers.py — shadow recomputation layer: pure-function delegators to shared interfaces, used for output-identity verification |
| Tracer | paper_trading/tracer.py — structured JSONL event tracing of every decision cycle + shadow comparison (signal, sizing, PnL) |
| Diagnostics | paper_trading/diagnostics.py — root-cause analysis layer: signal divergence classification, model distribution tracking, one-at-a-time feature impact scoring, volatility regime context, PnL decomposition |
| Shadow Memory | paper_trading/shadow_memory.py — persistent longitudinal store partitioned by asset+date (JSONL); immutable baseline with histogram-based model proba, signal mismatch, PnL error, and regime distributions |
| Drift Scoring | paper_trading/drift_scoring.py — 5-dimensional drift engine: model KL divergence, signal mismatch rate, PnL MAE, feature Jaccard stability, regime consistency; get_shadow_intelligence() produces per-asset drift report |
| Risk Governance | paper_trading/risk_governance.py — advisory risk layer: weighted composite score → LOW/MEDIUM/HIGH with risk flags, explanations, and non-binding recommended action |
| Shadow Actions | paper_trading/shadow_actions.py — non-binding execution advisor: action_type, exposure_adjustment, recommended_guardrails |
| Shadow Feedback | paper_trading/shadow_feedback.py — persistent behavioral dataset generator: append-only feedback events partitioned by asset+month |
| Shadow Analytics | paper_trading/shadow_analytics.py — offline aggregation: learning profiles, stability ranking, systemic pattern detection |
| Shadow Learning | paper_trading/shadow_learning.py — offline knowledge distillation: learning profiles, latent patterns, regime map, insights |
| Module | Description |
|---|---|
backtests/sandbox_retrain.py |
Isolated training pipeline with frozen DataLock audit trail |
backtests/model_comparator.py |
4-lens evaluation: model AUC/logloss, signal agreement/flip rate, portfolio PnL/drawdown, shadow entropy — all regime-stratified |
backtests/forward_test.py |
Walk-forward hold-out: trains fresh model on pre-cutoff data, computes forward Sharpe/hit rate/stability for baseline + candidate |
backtests/mas.py |
Model Acceptance Score: 6-dim weighted compression (model/signal/portfolio/shadow/forward/stress) with 4 hard gates |
backtests/model_evolution.py |
Temporal trajectory: append-only JSONL, MAS velocity, equilibrium bands, cross-asset convergence |
backtests/model_promotion_engine.py |
4-condition admission gate (performance/stability/consistency/safety) producing structured LIVE_CANDIDATE/PAPER_ONLY/SHADOW_ONLY/REJECT decisions |
backtests/adversarial_manifold.py |
9-regime perturbation engine across 4 axes (volatility, correlation, trend, noise); CMSS score + attractor drift + stability class |
- Feature augmentation is asset-dependent: 4-feature models suffice for assets with strong data-generating processes (GC: real yield + breakeven; NZDJPY: VIX + 10y spread). Augmenting with additional orthogonal signals (VIX delta, DXY momentum, shorter/longer momentum windows) lifted EURAUD from 0.57→2.28, CADJPY from 0.08→1.70, and USDCAD from 1.48→2.86 avg Sharpe in 5-year walk-forward. GC and NZDJPY regressed under augmentation — when the base features already capture the regime, more features add noise.
- Asset-specific driver features are mandatory: generic macro features fail on 28/30 assets tested; pair-specific features (VIX + bilateral yield spread for NZDJPY) flipped 0/7 → 5/7 positive windows
- Genuine diversification across 6 driver clusters: 6-asset portfolio spans momentum_crypto, real_asset, eur_cross, carry_fx, oil_carry, and usd_macro with max |r| < 0.40
- Feature interference is a real failure mode: macro features drowned by 25 price features until protected weight architecture separated them
- Macro features describe environment, not price response: yield_slope and real_yield_10y removed from XLF model because they stayed bearish through 2023–2024 rally; 2y_yield_delta_63 (direction, not level) was the fix
- EURUSD blocked at daily frequency: 8 years walk-forward showed 1.65% CAGR; requires COT positioning data
flowchart TD
A1[yfinance OHLCV] --> F
A2[FRED macro series] --> F
A3[Parquet cache] --> F
F[Driver atlas — asset-specific feature sets]
F --> |carry_fx| N1[NZDJPY: VIX + bilateral yield spread]
F --> |real_asset| N2[GC=F: real yield + breakeven + DXY]
F --> |momentum_crypto| N3[BTC: rate diff + yield delta + VIX + DXY + mom_10/21/63 + vs_spy_21/63]
F --> |eur_cross| N4[EURAUD: rate diff + DXY + VIX + delta + mom_21/63]
F --> |oil_carry| N5[CADJPY: VIX + spread + us_jp/ca_jp spread + DXY + mom_10/21/63]
F --> |usd_macro| N6[USDCAD: rate diff + DXY + VIX + delta + mom_21/63]
N1 & N2 & N3 & N4 & N5 & N6 --> L{tb20 vs fwd60}
L --> |tb20| TB[Triple barrier\npt_sl=2.0 · 20-bar timeout]
L --> |fwd60| FW[60-day forward return\n±2% threshold]
TB & FW --> M1[Regime classifier\nTREND/RANGE/VOLATILE/NEUTRAL]
TB & FW --> M2[Bootstrap deployment gate\np<0.20 for fwd60 · p<0.10 for tb20]
TB & FW --> M3[Walk-forward engine\n5yr train · 1yr test]
M1 & M2 & M3 --> S[Signal generator\nprob threshold 0.45]
S --> R[Risk engine\nvol-scaled sizing · halt conditions · prob_drift]
R --> P1[BTC 20%\nhalt DD −15%]
R --> P2[GC=F 20%\nhalt DD −8%]
R --> P3[EURAUD 22%\nhalt DD −8%]
R --> P4[NZDJPY 15%\nhalt DD −6%]
R --> P5[CADJPY 13%\nhalt DD −8%]
R --> P6[USDCAD 10%\nhalt DD −8%]
P1 & P2 & P3 & P4 & P5 & P6 --> Mon[Monitoring\nvalidity state machine · drift · dashboard]
QuantForge/
├── backtests/ # Sandbox model governance stack (Phases 11–14)
│ ├── sandbox_retrain.py # Orchestrator: trains isolated models + runs full eval pipeline
│ ├── model_comparator.py # 4-lens evaluation engine (model/signal/portfolio/shadow)
│ ├── forward_test.py # Walk-forward hold-out (Sharpe, hit rate, drawdown, stability)
│ ├── mas.py # Model Acceptance Score — 6-dim weighted compression + hard gates
│ ├── model_evolution.py # Trajectory tracking: velocity, equilibrium bands, convergence
│ ├── model_promotion_engine.py # 4-condition admission gate with structured failure modes
│ └── adversarial_manifold.py # 9-regime perturbation engine + CMSS stability scoring
├── paper_trading/ # Live engine + stdlib HTTP dashboard
│ ├── serve.py # stdlib HTTP server
│ ├── engine.py # Paper trading engine
│ ├── state_store.py # Versioned persistence boundary (EngineSnapshot, cache, journal)
│ ├── decision.py # TradeDecision / PositionIntent pure dataclasses
│ ├── position_manager.py # Position lifecycle (open, close, SL/TP, PnL)
│ ├── monitor.py # Entry point (data → signal → trade loop)
│ ├── wrappers.py # Shadow recomputation layer (pure-function delegators)
│ ├── tracer.py # Structured JSONL event tracing + shadow comparison
│ ├── diagnostics.py # Root-cause analysis: signal divergence, model dist, feature impact, regime, PnL decomp
│ ├── shadow_memory.py # Persistent longitudinal store, partitioned by asset+date, baseline builder
│ ├── drift_scoring.py # 5-dim drift engine: model KL, signal mismatch, PnL MAE, feature stability, regime consistency
│ ├── risk_governance.py # Advisory risk layer: weighted composite score + non-binding recommendations
│ ├── shadow_actions.py # Non-binding execution advisor: action_type, guardrails from drift+risk
│ ├── shadow_feedback.py # Persistent behavioral dataset generator (append-only, partitioned by asset+month)
│ ├── shadow_analytics.py # Offline aggregation: learning profiles, stability ranking, systemic pattern detection
│ ├── shadow_learning.py # Offline knowledge distillation: learning profiles, latent patterns, regime map, insights
│ ├── frontend/ # Dashboard UI (index.html, script.js, style.css)
│ └── models/ # 6 serialised XGBoost model pickles
├── shared/ # Strategy interfaces and registry
│ ├── __init__.py
│ ├── registry.py # StrategyRegistry singleton — per-asset routing, auto-register, validation
│ ├── model.py # ModelInterface ABC + XGBoostModel wrapper
│ ├── signal.py # SignalStrategy ABC + FixedThresholdStrategy
│ ├── sizing.py # PositionSizingStrategy ABC + VolTargetSizing
│ ├── pnl.py # PnLStrategy ABC + DefaultPnLStrategy
│ └── features.py # FeaturePipeline ABC + DefaultFeaturePipeline
├── scripts/ # Training & validation runners
│ ├── walk_forward_all.py
│ ├── train_all_assets.py
│ ├── gc_walk_forward.py
│ └── ...
├── equity/ # Walk-forward research scripts
├── archive/ # Retired model source code
├── models/ # Research models
│ ├── regime/ # Regime classifier
│ ├── ensemble/ # Model router
│ ├── mean_reversion/ # RSI + Bollinger for RANGE
│ ├── trend/ # Trend-following models
│ ├── volatility/ # Volatility models
│ ├── macro_expert_head.py
│ └── hybrid_ensemble.py
├── features/ # FeatureContract system + feature engineering pipeline
│ ├── contract.py # FeatureContract dataclass + runtime validators
│ ├── registry.py # FEATURE_REGISTRY — canonical contracts for 6 portfolio assets
│ └── builder.py # Generic builder — compute_macro_derived, build_features, compute_label, model_path
├── labels/ # Triple-barrier & meta-labeling
├── signals/ # Signal filtering, thresholding, generation, paper signal adapter
├── risk/ # Position sizing, drawdown controls, exposure limits
├── monitoring/ # Validity state machine, drift detection, dashboard backend, MLflow, weekly reports
├── data/
│ ├── loaders/ # yfinance, FRED, COT downloaders
│ ├── raw/ # Raw OHLCV parquet files
│ ├── processed/ # Feature-engineered datasets & walk-forward results
│ ├── live/ # Runtime engine state, equity history, logs, yfinance cache
│ └── weekly_pipeline.py
├── diagnostics/ # Model audits, sweeps, SHAP analysis, isolation tests
├── portfolio/ # HRP allocator, risk parity, correlation clusters
├── execution/ # Broker interface, order manager, portfolio sync, PaperBroker (simulated broker)
├── configs/ # YAML configs (paper trading, forex) + driver atlas
├── tests/ # Pytest test suite (148 tests, regression-verified across all 14 phases + feature augmentation)
├── docs/ # Project documentation, runbook, system overview
│ └── adr/ # Architecture Decision Records (ADR-000 through ADR-017)
├── adr/ # Additional ADRs (ADR-011 known issues)
├── notebooks/ # (placeholder — no notebooks yet)
├── .github/
│ └── workflows/ # CI pipeline (py_compile lint + pytest)
├── quantforge/ # Package root (version, logging)
├── main.py # Minimal entry point
├── monitor_all # Launch script (paper trading)
├── Makefile # Dev targets (install, test, lint, run, clean)
├── pyproject.toml # Project metadata & deps
└── requirements.txt # Pinned dependencies
Project documentation and architecture decisions live alongside the code:
| Path | Description |
|---|---|
docs/ |
Project documentation — guides, references, deep-dives |
docs/adr/ |
Architecture Decision Records — key design decisions and their rationale (ADR-000 through ADR-017) |
adr/ |
Supplementary ADRs — known issues and deferred decisions |
ADR entries follow the standard Michael Nygard template and are numbered sequentially.
git clone git@github.com:manuelhorvey/QuantForge.git
cd QuantForge
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install pytest pytest-cov # dev dependencies
export PYTHONPATH=$PYTHONPATH:.# Train models for all 6 portfolio assets (contract-driven pipeline)
python scripts/train_all_assets.py
# Run walk-forward validation
python scripts/walk_forward_all.py
# Run tests
make test
# Start live paper trading + dashboard
./monitor_allThe paper-trading engine has undergone a structural refactoring across 15 phases, each preserving byte-identical outputs:
| Phase | Description | Status |
|---|---|---|
| 1 | Tracer + shadow comparison + baseline snapshot + live contract + dependency guard | ✓ |
| 2 | Shared interfaces (model, signal, sizing, PnL, features) + AssetEngine migration | ✓ |
| 3 | Strategy registry with per-asset routing + research_mode flag | ✓ |
| 4 | Wrapper consolidation (delegate to shared interfaces) + PnL shadow + dead code removal | ✓ |
| 5 | Shadow diagnostics layer: signal divergence, model distribution, feature impact, regime context, PnL decomposition | ✓ |
| 6 | Persistent shadow memory (asset+date partitioned JSONL) + 5-dim drift scoring engine + baseline initialization | ✓ |
| 7 | Risk governance layer: weighted composite risk score, advisory exposure_multiplier, non-binding recommended action, /risk.json dashboard endpoints |
✓ |
| 8 | Shadow action layer: non-binding execution advisor with action_type (NONE/MONITOR/REDUCE/PAUSE), reason_codes, guardrails, /shadow-actions endpoints |
✓ |
| 9 | Shadow feedback loop: persistent behavioral dataset generator with derived metrics (agreement_score, instability_index, risk_alignment); offline analytics toolkit | ✓ |
| 10 | Shadow learning compilation: offline knowledge distillation — 5-dim learning profiles, latent pattern mining, regime behavior map, shadow insights | ✓ |
| 11 | Sandbox retraining harness: physically isolated model training + 4-lens evaluation engine (model/signal/portfolio/shadow comparison) with regime-stratified signal agreement and frozen DataLock audit trail | ✓ |
| 11.5 | Walk-forward test + M_stress: out-of-sample hold-out evaluation (Sharpe, hit rate, drawdown, stability); regime stress robustness sub-score calibrated to 0.5 at identity | ✓ |
| 12 | Model evolution tracking: per-asset trajectory JSONL with MAS velocity (Δ/retrain), acceleration, sub-axis drift, equilibrium band estimation, cross-asset convergence tracking | ✓ |
| 13 | Model promotion protocol: 4-condition multivariate admission gate (performance/stability/consistency/safety) producing structured decisions (LIVE_CANDIDATE/PAPER_TRADING_ONLY/SHADOW_ONLY/REJECT) with explicit failure modes | ✓ |
| 14 | Cross-manifold adversarial testing: 9-regime perturbation engine (volatility shock/compression, correlation break/inversion, trend flip/burst, noise inject/spike); CMSS stability score + attractor drift + stability class (ROBUST/MODERATE/BRITTLE) | ✓ |
| 15 | Feature augmentation (May 2026): Extended BTC, CADJPY, EURAUD, USDCAD with VIX delta, DXY momentum, shorter/longer momentum windows, and additional macro filters. CADJPY switched from fwd60 to tb20 labels. 5-year walk-forward showed avg Sharpe improvements of +0.44 (BTC), +1.62 (CADJPY), +1.71 (EURAUD), +1.38 (USDCAD). GC and NZDJPY left unchanged. Result: first LIVE_CANDIDATE promotion (EURAUD). | ✓ |
sandbox retrain
→ 4-lens comparison (model/signal/portfolio/shadow)
→ walk-forward test (6mo hold-out)
→ MAS (6-dim weighted compression: model/signal/portfolio/shadow/forward/stress)
→ adversarial manifold (9 regime perturbations → CMSS)
→ trajectory append (velocity + equilibrium bands)
→ promotion gate (4-condition admission) → structured decision
The live engine is swap-ready but locked. The sandbox is a proper model selection energy landscape: stable models sit in an equilibrium basin (~80 MAS), genuine structural improvements must climb past 88 with >0.6 stress robustness, and degraded models fall below 70. No single metric can force deployment — MAS is a manifold position, not a pass/fail score.
After May 2026 feature augmentation, the promotion gate produced its first LIVE_CANDIDATE (EURAUD: MAS=85.58, M_stress=0.63, 4/4 conditions met) and PAPER_TRADING_ONLY (CADJPY: MAS=85.22, 3/4 conditions, M_stress=0.59 — 0.0064 below deployment threshold).
- EURAUD shadow live test (30-day) — first LIVE_CANDIDATE promotion
- Live broker integration (Alpaca / Interactive Brokers) — PaperBroker(BrokerInterface) contract exists
- Adversarial manifold expansion: bootstrap volatility shocks, permuted regime segments, synthetic spread widening
- Shadow intelligence dashboard integration — surface drift, CMSS, and MAS trajectory in web UI
- CADJPY re-evaluation — 0.0064 below M_stress threshold; monitor if next retrain cycle crosses 0.60
- AUDJPY re-evaluation post-November — model trained, deferred pending correlation analysis
- Ensemble manifold synthesis — merge accepted models via regime-conditional weighting
- External regime validation sets — periodic "unknown regime injection" tests to prevent over-calibration
- EURUSD/GBPUSD with COT data pipeline
Assets pending validation: ETH-USD (momentum_crypto cluster), XLU/XLRE (yield_equity cluster). AUDJPY (model trained, deferred post-November — r=0.87 with NZDJPY).
Blocked pending data acquisition: EURUSD, GBPUSD (need CFTC COT weekly positioning data).
- Paper trading only (no real capital at risk)
- Weekend data staleness for equities/FX
- No live broker execution — PaperBroker(BrokerInterface) provides the contract; Alpaca/IBKR stubs exist
- EURUSD and GBPUSD blocked — requires COT positioning data
- Model validity depends on PSI distribution stability
- Paper trading uses XGBoost per asset with 4–10 features (asset-dependent); research ensemble (
HybridRegimeEnsemble) not yet deployed — registry is swap-ready but locked (a singleStrategyRegistry.register_model()call would activate it) - Yahoo Finance data can be rate-limited; engine mitigates with exponential backoff (5s/15s/45s) + parquet cache fallback
This project is for research and educational purposes only. It is not financial advice. Trading involves substantial risk of loss. Past performance does not guarantee future results.
MktOwl
Focus: Macro-driven systematic trading • Walk-forward validation • Production-grade research engineering
Contributions, issues, and suggestions are welcome.