THE ALGORITHM

The algorithm.

The fund is a vehicle. The product is the algorithm — a proprietary machine-learning framework designed, trained, and supervised in-house. Every position taken by the fund is produced by the same system, bounded by human-defined risk, executed under continuous supervision.

ARCHITECTURE

Learned where it helps. Hard-coded where it matters.

A hybrid architecture combining deep-learning sequence models, gradient-boosted ensembles, and hard-coded rules-based filters. Trained on eight years of live digital asset market data, deployed under continuous supervision.

Sequence Models

Deep-learning sequence models trained on multi-timeframe market data, identifying patterns human analysis cannot hold in working memory across years of price action.

Feature Engineering

Hundreds of engineered features spanning price, volume, volatility, and on-chain signals, refined through continuous backtesting against live market regimes.

Feature Engineering

Hundreds of engineered features spanning price, volume, volatility, and on-chain signals, refined through continuous backtesting against live market regimes.

Ensemble Layer

Gradient-boosted ensembles combine independent model outputs into a single confidence-weighted signal, reducing reliance on any one model’s failure mode.

Ensemble Layer

Gradient-boosted ensembles combine independent model outputs into a single confidence-weighted signal, reducing reliance on any one model’s failure mode.

Rules Boundary

Hard-coded rules-based filters sit outside the learned layers, enforcing constraints no model is permitted to override regardless of its confidence.

Rules Boundary

Hard-coded rules-based filters sit outside the learned layers, enforcing constraints no model is permitted to override regardless of its confidence.

PROCESS

Signal to execution.

The algorithm operates as four sequential layers. Each has a defined role. Each has boundaries the layer above cannot cross.

1 Signals

An ensemble of sequence models and gradient-boosted trees produces a confidence-weighted signal output for each market and timeframe under review.

Signals

An ensemble of sequence models and gradient-boosted trees produces a confidence-weighted signal output for each market and timeframe under review.

2 Execution

Latency-aware execution passes every signal through a risk layer gate before it reaches the market. There is no exception path around this gate.

Execution

Latency-aware execution passes every signal through a risk layer gate before it reaches the market. There is no exception path around this gate.

3 Risk

Hard-coded caps trigger automatic de-risking when breached. Recovery from a de-risked state requires principal review, not an automated reset.

Risk

Hard-coded caps trigger automatic de-risking when breached. Recovery from a de-risked state requires principal review, not an automated reset.

4 Adaptation

Rolling retraining is validated in shadow mode against live markets before promotion. Every model version is recorded in an auditable history.

Adaptation

Rolling retraining is validated in shadow mode against live markets before promotion. Every model version is recorded in an auditable history.

TRACK RECORD

Proprietary · 2020–2026

Figure committed capital

Profit

Average Yearly Return

Worst Peak to trough

Maximum drawdown

Past performance is not indicative of future results. Track record reflects live systematic execution on principal capital; audited third-party performance records provided to qualified investors under NDA.

Verification before allocation.

Qualified investors receive audited third-party performance records under NDA, full methodology documentation, and direct access to the principals before any allocation discussion begins. We consider that sequence non-negotiable — in both directions.

Contact