Methodology

How Aksoy Capital's Darwin AI generates research insights — full transparency on architecture, data, validation, and limitations.

1. The Darwin Architecture

Aksoy Capital's research engine is a multi-layer evolutionary algorithm called Darwin AI. Each layer specializes in a different time horizon, asset class, or decision pattern. They evolve in parallel and their outputs are blended by a top-level allocator. The number of layers, their internal weighting, and their selection thresholds are proprietary.

Specific universes, layer counts, and weighting schemes are proprietary.

2. Genetic Algorithm Mechanics

Each layer is a genetic algorithm (GA). Populations of candidate strategies (chromosomes) compete across many generations using tournament selection with elitism, uniform crossover, and Gaussian mutation on continuous genes. Population sizes, generation counts, mutation rates, and tournament parameters are tuned per layer and are not published.

2.1 Chromosome Genes

Each chromosome encodes four broad gene categories:

1. Technical analysis genes — momentum, mean-reversion, volatility, and volume signal parameters, plus entry / exit thresholds and stop-loss / take-profit envelopes.
2. AI context genes — when and how heavily to weight the AI contextual signal, and the conditions under which it is consulted.
3. Universe distribution genes — index-level allocation across the covered universes.
4. Cash / risk gene — capital-preservation buffer.

Exact gene names, valid ranges, and the relationships between them are proprietary.

3. Training Data

4. Train/Validation/Test Split

• Training: 2005-2022 (18 years)
• Validation: 2023-2024 (2 years)
• Test: 2025 (out-of-sample, never used during training)

4.1 Walk-Forward Validation

To guard against overfitting, we use multi-year rolling training windows with non-overlapping validation periods. Each chromosome is tested across multiple distinct market regimes (low-volatility, high-volatility, drawdown, recovery) before being promoted to live deployment. Exact window lengths and step sizes are tuned per layer and proprietary.

5. Benchmarks

All Darwin layers are benchmarked against:

• SPY — passive S&P 500 buy-and-hold (the primary benchmark)
• 60/40 portfolio — 60% SPY + 40% IEF (10-year Treasury)
• 12-month momentum factor — academic factor model

Promotion rule: A new layer or chromosome must beat SPY on a Sharpe-ratio basis on the validation window before being deployed live.

6. Fitness Function

The fitness function rewards risk-adjusted return (Sharpe ratio) and penalizes drawdowns, transaction costs, AI query costs, and known biases in the training data (see §8). The relative weighting of each term, and any mode-specific extensions (e.g. win-rate bonus for the Trader mode, Calmar emphasis for the Investor mode), are proprietary and tuned per layer.

7. AI Context Layer

Darwin AI uses an advanced large language model as its contextual layer. To preserve methodological integrity during backtesting, every AI signal is pre-computed before the simulation starts and cached to disk; the simulator never queries the model live during a historical backtest. This eliminates any chance of look-ahead contamination from the model provider’s own training data.

In live deployment the system may escalate to a stronger model class under high-stress market conditions — elevated volatility, abnormal volume, fresh corporate-event filings, or central-bank / earnings-day windows. The exact thresholds, the specific model identifiers used at each tier, and the prompt set are not published.

8. Survivorship Bias Disclosure

9. What This Is NOT

• Not investment advice. Aksoy Capital is not a registered investment adviser (RIA) under SEC rules. Our research is published for educational purposes.
• Not personalized. Outputs do not consider your tax bracket, time horizon, or risk tolerance.
• Not a guarantee. Past simulation performance does not predict future returns. Real trading involves slippage, fees, taxes, and market impact not modeled in backtests.
• Not a replacement for due diligence. We strongly encourage users to read SEC filings, conduct fundamental research, and consult a licensed financial advisor before making investment decisions.

9.5 AI Autonomy Policy (L0-L5)

Aksoy Capital Darwin AI publishes algorithmically-generated research at scale. Because no human edits every artifact, we operate under a published autonomy policy that names which level (L0-L5) every channel runs at, and which gates apply.

Current commitment: All L4 (supervised autonomous) - human monitors dashboards, retracts on signal, but does not pre-approve every artifact. No channel runs at L5 (full autonomy). The publishing system cannot self-modify; every change to prompts, models, or schedules requires a human commit on the application repository.

• QA gates at L4: Forbidden terms, numeric consistency, attribution check, ticker whitelist, HTML sanitizer, crosslink budget, and the Aksoy Capital Darwin AI transparency badge on every report.
• Retraction: Material errors are corrected within 1 hour (URL preserved, body replaced with retraction notice), follow-up note in next morning brief, and a new QA gate added within 7 days.
• Editorial owner: Ali Aksoy is the named editor-in-chief for every L4 channel.

Full policy text including the channel-to-level mapping table and quarterly review process: /docs/ai-policy/ (last reviewed 2026-05-06).

10. Updates and Versioning

Each Darwin layer has a versioned model checkpoint. When we retrain or replace a model, we publish:

• Model version + retrain date
• Walk-forward Sharpe vs SPY (validation set)
• What changed (new genes, new training window, etc.)

You can subscribe to model-update notifications via the dashboard.

11. Reproducibility

While we do not publish the exact production model weights, all of the following are open and inspectable:

• Training data sources (above)
• Algorithm architecture (this page)
• Validation methodology
• Daily/weekly research outputs (free tier)

12. Questions?

Email [email protected] with subject "Methodology" — we respond to specific questions about our approach within 5 business days.