The AI desk

A complex system that trades one book, on your behalf.

NeuralCryptoCapital runs its own trading program: a stack of statistical models, reinforcement learning and hard risk controls that decides when the pool buys, swaps and sells. Clients never place an order — the capital is managed for them.

Signal→
Sizing→
Pre-trade limits→
Execution→
Post-trade audit

Five gates sit between a signal and the book. A decision that fails any gate is dropped and logged — the desk would rather miss a trade than carry an uncontrolled one.

Four systems, one decision.

01

Ingestion & feature store

Order-book deltas, trades, funding rates and on-chain flows are normalised into one time-aligned store. Each signal is timestamped at the source, so nothing in training or replay can see the future.

Streaming ingest · point-in-time joins · Parquet snapshots

02

Regime detection

A hidden Markov model classifies the market as quiet, trending or stressed. Model weights, position limits and holding horizons all shift with the regime instead of assuming one market forever.

Gaussian HMM · Bayesian change-point detection

03

Forecasting & policy

Transformer sequence models score the next 1–72 hours across assets. A reinforcement-learning policy converts those scores into orders, with slippage and inventory cost inside the reward.

Attention models · PPO execution policy · reward shaping

04

Risk & settlement

Every order clears pre-trade limits, then a post-trade audit records the decision, the inputs that produced it and the realised outcome. Nothing reaches the book without passing both.

CVaR budget · fractional-Kelly caps · immutable decision log

Mathematics

What is actually being computed.

No single model decides alone. Each layer answers a different question, and the output of one becomes a constraint on the next.

Transformer attention

Multi-head self-attention over price, book and on-chain sequences gives long memory without hand-built indicators, and scales across many assets at once.

Bayesian state-space models

Kalman filtering estimates a latent fair value from noisy venue quotes, keeping execution anchored when prices disagree across exchanges.

Hidden Markov models

Latent market states are inferred probabilistically, letting the desk distinguish trend from chop before it adds size.

Deep reinforcement learning

Proximal policy optimisation learns when to work an order and when to wait, penalised for slippage, fees and leftover inventory.

Convex optimisation

Allocation is solved as a constrained quadratic program: risk budget, per-asset caps and liquidity limits are hard constraints, not preferences.

Stochastic control & Kelly

Position size follows expected edge over variance, then a fractional-Kelly rule and a CVaR ceiling decide how much of that edge is allowed on.

Monte Carlo stress testing

Thousands of synthetic paths run daily against the live policy to test drawdown behaviour under shocks history has not shown yet.

Walk-forward validation

Models are refitted on purged, embargoed folds and promoted only if they hold up out of sample, so reported skill is not leakage.

Generative systems

Where language models earn their place.

Research agents with retrieval

Language-model agents read exchange notices, filings and market news, retrieve the relevant passages and emit structured event tags. Any tag that would change exposure is reviewed by a human before it reaches the policy.

Scenario generation

Generative models write plausible stress narratives — an exchange outage, a stablecoin de-peg, a liquidation cascade — which are compiled into new Monte Carlo scenarios for the risk budget.

Adversarial model review

Before promotion, each model is attacked by generated counter-examples and synthetic edge cases; failures become regression tests in the release gate.

Human in the loop

Generative output never places an order. It can only change the context the policy sees, and every promoted change leaves a signed audit trail.

Guardrails are part of the system, not a footnote.

Hard limits

Per-asset caps, gross exposure and a daily loss budget are enforced before any order leaves the desk.

Independent review

Models are reviewed by the risk desk before promotion, and promoted changes are versioned and auditable.

Kill switch

The book can be flattened and the policy halted manually at any time, independent of the models.

This page describes the design of the trading program. Performance figures shown across this site are simulated for demonstration and are not audited results, confirmed client returns or a forecast. Nothing here is investment advice.

Let the desk work on your capital.

Open an account from $10 and track the pool daily.

Start investing