Key Takeaways
- • AllocFlow, a reinforcement-learning model, posted a 51.38% annualized net return in backtests on hourly crypto perpetual futures, up from 27.79% after its initial imitation training stage.
- • The model's Calmar ratio, comparing annualized return to maximum drawdown, rose from 0.955 to 4.043 over the two training stages.
- • Results include trading fees, and the same trained model recorded a 4.90 Calmar ratio on six months of unseen 2026 data without retraining.
- • All figures are historical backtests, not live trading results, and the source states they do not establish how AllocFlow would perform in live markets.
Summary
AllocFlow converts hourly market histories into target portfolio weights, where positive weights represent long positions and negative weights represent short positions. Rather than forecasting each asset separately and then constructing a portfolio from those forecasts, the model evaluates long and short positions jointly under shared capital constraints.
Training occurred in two stages. The model first learned to imitate an existing trading strategy to establish 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, rewarding results at the portfolio level. The main test used perpetual futures, which are contracts without an expiration date that can be used for long or short positions. The reported returns are net of trading fees, a relevant consideration for a strategy that adjusts positions frequently.
The same trained model was also evaluated on six months of unseen 2026 data without retraining, where it recorded a 4.90 Calmar ratio. The article stresses that both sets of figures describe backtests rather than trades placed in a live market, and that historical results under these test conditions do not establish how AllocFlow would perform in live trading.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.