AllocFlow Raises Backtested Crypto Futures Return to 51.38% With AI
Reinforcement learning lifted the model’s annualized net return from 27.79% after imitation training. The results come from historical tests, not live trading.

Researchers’ AllocFlow model posted a 51.38% annualized net return in historical cryptocurrency perpetual futures tests after reinforcement learning, testing an AI approach that chooses long and short positions across assets together.
That return was up from 27.79% after the model’s initial imitation training. Its Calmar ratio, which compares annualized return with maximum drawdown, rose from 0.955 to 4.043. Both figures describe backtests rather than trades placed in a live market.
AllocFlow turns hourly market histories into target weights for a portfolio. A positive weight represents a long position, while a negative weight represents a short position. The model considers those positions together under shared capital constraints instead of making a separate forecast for each asset and then building a portfolio from the forecasts.
Training took two stages. First, the model learned to imitate an existing trading strategy, giving it a starting set of portfolio choices. Reinforcement learning then refined those choices by comparing eight candidate portfolios at a time against the following 24 hourly market bars. That second stage rewarded results for the portfolio as a whole.
The main test used hourly cryptocurrency perpetual futures. Perpetual futures are contracts without an expiration date that can be used for long or short positions.
The reported net returns included trading fees. That matters because a strategy that frequently adjusts positions can lose part of its gains to trading costs.
The same trained model was also tested on six months of unseen 2026 data. It recorded a 4.90 Calmar ratio during that period without retraining. Historical results under these test conditions do not establish how AllocFlow would perform in live trading.