AlphaFactory: An Autonomous Research Engine for Systematic Trading

The Problem

Systematic strategy research is a grind. A single hypothesis — from idea through backtest, out-of-sample validation, platform parity check, and paper trading — can take a skilled quant 2–4 weeks of focused work. Most hypotheses fail. The survival rate from idea to production-ready strategy is typically under 1%.

This creates a brutal math problem: if you need 10 production strategies to build a diversified portfolio, and 99% of ideas fail, you need to rigorously evaluate roughly 1,000 hypotheses. At 2 weeks each, that's 40 years of work.

The Solution

AlphaFactory inverts this bottleneck by treating strategy research as an automated pipeline. An LLM-driven research loop continuously generates hypotheses, writes backtest code, runs walk-forward validation, reviews results for statistical artifacts, and queues survivors for independent verification on a separate execution platform.

The system operates 24/7 with minimal human intervention. My role shifts from doing the research to directing the research — setting constraints, reviewing edge cases, and making final deployment decisions.

Core Principles

• No curve-fitting. Every strategy must survive walk-forward validation across multiple out-of-sample periods, not just look good on in-sample data.
• Platform independence. Python backtests are cross-validated on an independent execution platform to eliminate backtesting engine discrepancies. A strategy that only works in one environment isn't a real strategy.
• Fail fast. The pipeline is designed to kill bad ideas quickly and cheaply. Most hypotheses are eliminated in seconds, not days.

Architecture

The platform is built on four layers:

Hypothesis Generation

The research engine draws from multiple sources to generate strategy hypotheses:

• Academic literature: Calendar effects, mean reversion anomalies, volatility regime patterns
• Market microstructure: Opening range dynamics, session transitions, institutional flow patterns
• Indicator combinations: Systematic exploration of signal confluence across technical indicators
• Cross-market analysis: Structural relationships between correlated instruments

Each hypothesis is documented with a clear rationale before any code is written. "What structural behavior would cause this edge to exist?" is the first question, not "Does the backtest look good?"

Multi-Tier Validation

Strategies pass through a rigorous multi-stage validation pipeline, with each stage designed to catch a specific class of false positive:

Stage	Purpose	Kill Rate
Quick Screen	Vectorized scan, eliminates obvious losers in <5 seconds	~70%
Walk-Forward Backtest	Event-driven simulation with proper execution modeling. 5+ folds, 5+ years of data.	~20%
Lookahead Bias Review	Automated code review checking for future data leakage, same-bar execution, and date logic errors	~3%
Independent Platform	Cross-validation on a separate bar-by-bar execution engine with realistic commission and slippage	~5%
Paper Trading	Live market execution with real data feeds, no capital at risk	~1%

The cumulative survival rate from hypothesis to production is approximately 0.7% — 12 validated strategies from over 1,700 tested.

Walk-Forward Protocol

Every strategy is tested using expanding-window walk-forward validation:

• Training window: 2–3 years of historical data
• Test window: 6–12 months of completely unseen data
• Minimum folds: 5 non-overlapping out-of-sample periods
• Consistency requirement: >60% of folds must be independently profitable

Platform Parity

An early and expensive lesson: Python backtests can inflate performance by 30–200% compared to bar-by-bar execution engines due to execution timing assumptions. A signal generated at bar N's close that enters at the same close (rather than the next bar's open) creates a systematic lookahead bias that flatters results.

The solution was migrating to an event-driven backtesting framework that matches the production execution engine within 0.3%, then requiring independent confirmation with realistic transaction costs ($5.00 round-trip commission + 1 tick slippage per trade).

Execution Infrastructure

Validated strategies are translated into a domain-specific strategy language and deployed on a production signal platform:

• Signal source: Strategies encoded in a domain-specific language, running on live market data
• Relay: Dockerized FastAPI service on a cloud VPS with auto-TLS
• Broker integration: REST API clients for futures brokers with multi-account fanout
• Position tracking: Real-time state machine with P&L monitoring and alerting
• Risk controls: End-of-day position flattening, duplicate signal detection, maximum position limits

The relay processes signals in <100ms and supports simultaneous execution across multiple funded accounts from a single signal source.

Autonomous Research Loop

The research engine runs continuously:

1. Generate — LLM produces strategy hypotheses from a prioritized research queue
2. Test — Hypotheses are automatically backtested with walk-forward validation
3. Review — Results are checked for statistical artifacts and lookahead bias
4. Synthesize — Cross-experiment patterns are analyzed to inform future hypothesis generation
5. Queue — Survivors are queued for independent platform validation

The system uses a multi-armed bandit approach (UCB1) to balance exploring new strategy families against exploiting promising directions. Failed experiments feed back into the hypothesis generator — understanding why something doesn't work is as valuable as finding what does.

Results

By the Numbers

Metric	Value
Total hypotheses evaluated	1,700+
Strategies passing walk-forward validation	~50
Strategies confirmed on independent platform	12
Survival rate (hypothesis → production)	0.7%
Markets covered	NQ (Nasdaq 100), ES (S&P 500) equity index futures
Strategy types	Calendar effects, breakout, mean reversion, trend following
Portfolio trade frequency	~200 trades/year across all strategies
Validation data	5+ years of 1-minute futures data (2008–2026)

Portfolio Characteristics

The validated portfolio is intentionally diversified across:

• Time horizons: Intraday (minutes), swing (hours), position (days)
• Market conditions: Strategies that perform in trending markets are paired with those that perform in mean-reverting or range-bound conditions
• Instruments: Primarily NQ (higher volatility, momentum-driven) with ES (lower volatility, mean-reverting) for diversification
• Session windows: Different strategies target different parts of the trading day, reducing signal correlation

No single strategy contributes more than 25% of expected annual P&L, and no two strategies share the same entry logic.

What Didn't Work

Intellectual honesty requires acknowledging the dead ends:

• Precious metals (Gold, Silver): 13 strategy families tested, zero validated. The microstructure is fundamentally different from equity indices.
• Crude oil: 30+ strategies tested, one marginally validated. Transaction costs ($42.80 round-trip) kill most intraday edges.
• ML-based approaches: XGBoost ensembles, regime classifiers, and neural network signal generators all showed in-sample promise that evaporated out-of-sample.
• Complex indicator combinations: More indicators does not equal better performance. The best-performing strategies use 1–3 simple filters. Complexity is the enemy of robustness.

Capital Scaling Model

The strategies are deployed through funded trading accounts provided by proprietary trading firms, which offer leverage without personal capital risk:

• Evaluation: Pass a simulated trading evaluation demonstrating consistent profitability and risk management
• Funded accounts: Receive a funded account with real capital (typically $50,000–$150,000)
• Profit split: Retain 80–90% of trading profits
• Scaling: Multiple funded accounts across multiple firms, each running the same validated strategies

This model allows scaling capital deployment without proportional capital requirements. The relay infrastructure supports simultaneous execution across all accounts from a single signal source.

Risk Management

• All strategies flatten positions by end of day (no overnight futures exposure)
• Maximum drawdown limits enforced at both the strategy and account level
• Position sizing calibrated to account risk parameters
• Automated monitoring with real-time alerts for anomalous behavior

Technology Stack

Component	Technology
Research orchestration	Python, LLM agents (Anthropic Claude)
Backtesting	Event-driven framework, custom walk-forward harness
Cross-validation	Independent execution platform, automated via browser automation
Signal generation	Domain-specific strategy language on live market data
Execution relay	FastAPI, Docker, Caddy (auto-TLS)
Broker connectivity	Async REST API clients (Python)
Data	1-minute continuous futures, 2008–present
Monitoring	Webhook alerts, real-time dashboard
Infrastructure	Cloud VPS, dedicated research workstation

About

Greg Kitzmiller is a systems architect with a background in high-scale cloud infrastructure at Microsoft. AlphaFactory represents the intersection of modern AI capabilities and quantitative trading — using large language models not to predict markets directly, but to dramatically accelerate the research process that identifies durable structural edges.

Contact: greg@thequietconviction.com