Strategy Workflow

Comprehensive strategy development workflow for quantitative trading, from hypothesis to validated production deployment.

Overview

This skill provides a complete framework for developing, testing, and validating trading strategies. It supports:

• - Hypothesis-driven strategy development
• Multi-GPU backtesting on Vast.ai
• Bayesian hyperparameter optimization with Optuna
• Walk-forward validation and out-of-sample testing
• Automated tearsheet generation

Entry Points

Control Plane (Swarm Orchestration)

Always-on watchdog loops that manage hardware utilization and self-healing:

CODEBLOCK0

For local environments, set explicit paths:

CODEBLOCK1

Work Plane (Parallel Execution)

Unified wrapper that starts control plane and launches parallel work:

CODEBLOCK2

Multi-GPU, multi-symbol execution:

CODEBLOCK3

Single-Symbol Pipeline

For focused optimization on a single asset:

CODEBLOCK4

Strategy Development

1. Hypothesis Formulation

Define your strategy hypothesis in measurable terms:

• - What market inefficiency are you exploiting?
• What is the expected holding period?
• What are the entry/exit conditions?
• What is the target risk-adjusted return?

2. Feature Selection

Identify relevant features for signal generation:

• - Price-based (OHLCV, returns, volatility)
• Technical indicators (EMA, RSI, Bollinger Bands)
• Multi-timeframe features (MTF resampling)
• Volume analysis (PVSRA, VWAP)
• Market microstructure (order flow, spread)

3. Signal Generation

Convert features into actionable signals:

• - Directional bias (trend following, mean reversion)
• Entry conditions (threshold crossings, pattern recognition)
• Exit conditions (take-profit, stop-loss, trailing stops)
• Position sizing rules

4. Position Sizing

Implement risk-aware position sizing:

• - Fixed fractional
• Kelly criterion
• Volatility-adjusted
• Regime-dependent scaling

Backtesting

Pre-Flight Validation

MANDATORY before every optimization run:

CODEBLOCK5

Validation checks:

• - Data >= 90 days with no gaps/NaN
• Min trades >= 30 for statistical significance
• MTF resampling implemented correctly
• No look-ahead bias

Multi-GPU Execution on Vast.ai

Deploy to cloud GPU instances for large-scale parameter sweeps:

CODEBLOCK6

Prescreening with Vectorized Backtests

Phase 0: GPU-accelerated parameter screening:

• - Generate N random parameter combinations
• Batch evaluate on GPU
• Filter by minimum trades (30+)
• Return top K by Sharpe ratio

Performance baseline (RTX 5090, 730d lookback, 250k combos): ~4s per mode.

Full Backtests with NautilusTrader

Phase 1: Event-driven backtesting for top candidates:

• - High-fidelity simulation with realistic execution
• Slippage and commission modeling
• Multi-asset portfolio backtests

Parameter Optimization

Optuna for Hyperparameter Search

Phase 2: Bayesian optimization with warm-start from prescreening:

CODEBLOCK7

Grid Search vs Bayesian Optimization

Pruning Strategies

• - MedianPruner: Prune if worse than median of completed trials
• PercentilePruner: Prune bottom X% of trials
• HyperbandPruner: Multi-fidelity optimization
• SuccessiveHalvingPruner: Aggressive early stopping

Distributed Optimization

For large-scale runs, use persistent storage:

CODEBLOCK8

Walk-Forward Validation

Rolling Window Validation

Slide the training/test window through time:

CODEBLOCK9

Parameters:

• - train_window: Training period length
• INLINECODE1: Out-of-sample test length
• INLINECODE2: Window advancement increment

Anchored Walk-Forward

Expand training window while sliding test window:

CODEBLOCK10

Use when historical regime diversity improves model robustness.

Epoch Selection Criteria

Intelligent selection of training periods:

• - Regime-aware: Match training regimes to expected deployment conditions
• Volatility-adjusted: Include both high and low volatility periods
• Event-inclusive: Ensure major market events are represented
• Recency-weighted: Emphasize recent data while maintaining diversity

Out-of-Sample Testing

Final validation phase:

• - Hold out 20-30% of data for final OOS test
• No parameter tuning on OOS data
• Monte Carlo stress testing
• Regime-conditional performance analysis

SLOs and Guardrails

Utilization Targets

• - CPU utilization target: >= 70%
• GPU utilization target: >= 70%
• No silent GPU fallback for GPU sweeps

Hardware Watchdog Hooks

Enforced by:

• - INLINECODE3
• INLINECODE4

Capacity Monitoring

Control plane loops monitor:

• - Worker health and liveness
• Progress artifact freshness
• Resource utilization
• Job queue depth

Self-healing actions:

• - Automatic worker restart on crash
• Fill lanes for underutilized resources
• Cooldown guardrails to prevent thrashing

Tearsheet Generation

Generate QuantStats-style performance reports:

CODEBLOCK11

See tearsheet-generator skill for detailed visualization options.

Multi-Provider Orchestration

PAL MCP Integration

Attach PAL as an MCP server for research/consensus across multiple model providers:

• - Config template: INLINECODE6
• Docs: INLINECODE7
• Providers: OpenRouter, OpenAI, Anthropic, xAI, local models

Resources

Documentation

• - VectorBT Documentation
• NautilusTrader Docs
• Optuna Documentation
• QuantStats

Project References

• - config/workflow_defaults.yaml - Default configuration
• INLINECODE9 - Model policy (advisory)
• INLINECODE10 - Detailed runbook
• INLINECODE11 - Hook contracts
• INLINECODE12 - VectorBT PRO integration

Results Structure

CODEBLOCK12