Running 24/7 · updated every 15 minutes

An AI that learns to trade,
and keeps evolving.

ASPI (Autonomous Self-evolving Profit Intelligence) is a reinforcement-learning trading system for crypto perpetual futures. It learns from years of market history, has to prove itself on markets it has never seen, and then runs by itself around the clock.

H1 return
—
 
H2 return
—
 
Closed trades
—
across both books
Days tracked
—
 

Performance

Cumulative return of each book as a percentage of its starting allocation, on closed trades, after trading fees and funding. The curve updates on its own as trades close.

Loading…

H1 · Concentrated

Three high-volatility markets, each traded by its own specialist model. Fewer, larger trades, with higher return potential and larger swings.

Return—
Max drawdown—
Trades—

H2 · Broad

Fifteen markets traded by models bred on a much wider universe. More trades and more diversification, so each one moves the book less.

Return—
Max drawdown—
Trades—

What ASPI is

Most trading bots follow rules a person wrote. ASPI writes its own. It is a neural network trained with reinforcement learning: it trades through years of market history, over and over, in a detailed simulator, is rewarded for good risk-adjusted outcomes and penalised for losses, and slowly learns a strategy no one programmed into it.

The "self-evolving" part sits one level higher. Between training rounds an evolutionary process adjusts what the system is rewarded for, keeping changes that make it better on data it was not trained on. The aim is a system that improves its own learning, not only its trades.

Finished models are tested hard before they ever trade, then frozen. New models join a growing library instead of overwriting old ones.

  • Learns, not scripted. Entries, stops, targets, sizing and exits all come from the trained policy.
  • Judged on unseen data. A model is only trusted on markets and months it never saw in training.
  • Built for real costs. Fees, funding, slippage and liquidation are part of the training world.
  • Runs by itself. The signal engine works 24/7 and recovers automatically after restarts.

How it works

From raw market data to a live order, in five stages.

Learn

A policy network trains in a market simulator built on years of 15-minute data, deciding when to trade, which way, how much risk to take and when to get out.

Evolve

An evolutionary tuner mutates the reward function between training cycles and keeps what generalises, so the training objective itself is optimised.

Certify

Candidates must pass a full month of market data they have never seen, over many randomised runs and on a mirror-image market, to show the edge is not just one trend.

Deploy

Certified models are frozen and join a library. Each market is matched to the model whose conditions suit it best. Models are never retrained.

Execute

A signal engine re-evaluates every model on each new bar, and an execution layer mirrors its target positions with passive limit orders inside strict risk limits.

What's inside

The main building blocks. Exact settings, inputs and the model zoo stay private.

Hierarchical policy

Separate heads decide whether to trade, how to shape the entry (stop and target), and whether to hold or close, each trained on its own question.

Distributional critic

Instead of one expected value, the model predicts a whole distribution of outcomes, so it can avoid trades with heavy downside tails.

Risk-normalised sizing

Every position is sized from its stop distance, so each trade risks a controlled share of the book whatever the volatility.

Commitment exits

Trades are held for committed periods rather than second-guessed every bar, which removes the noise-driven early exits that kill most strategies.

Rich market context

Multi-timeframe price structure, volatility regime, funding cycle, time of day and cross-market signals: far more than one chart.

Mirror training

Part of training runs on inverted markets, so the model cannot get away with simply being long in a rising market.

Passive execution

Orders rest on the book as maker limits, which keeps costs to a fraction of market orders and makes slippage measurable.

Safety rails

Daily loss limit, per-market notional caps, an instant kill switch that flattens everything, and alerting that watches the whole stack.

Pre-registered tests

What counts as success is written down before each month starts. Results are scored against that rule, never re-scored after the fact.

Principles

Trading research is full of strategies that look perfect in a backtest and fail the moment they meet a new month. ASPI is built around not fooling itself.