Why an AI trading on its own tends to lose money

The experiment

In 2025, nof1 had six language models trade crypto perpetual contracts with real money, with no one reviewing their decisions. Each model started with $10,000. In the first season of Alpha Arena, four of the six finished below their starting balance. The results are published at https://nof1.ai.

Why this happens

• Language models generate plausible text. Plausible does not mean correct, and in the market a confident mistake costs money.
• The same prompt can produce different decisions at different times, which makes it impossible to measure and fix the strategy.
• Without hard limits, a string of mistakes turns into a large loss before anyone notices.
• Simple math, like position size and stop distance, is prone to error when a text model writes it.

How Aval splits the work

In Aval, each step has a different owner. Code reads the market and calculates the indicators, the same way every time. JEV, a model trained to decide rather than to write, gives the score and can veto. Risk limits sit out of reach of any model: maximum daily loss, size per position, allowed pairs and pauses after stops.

The final decision is yours. You approve each signal or turn on auto mode from a minimum score you choose, always within the limits.

Measure before you trust

The best defense against promises is a scoreboard. The Aval account trades on paper, with real Hyperliquid prices and fees deducted, and the results are public, including how often each score range won. That way you can see whether the score works before putting real money in.