Forecasting a profitable asset and deciding how much of a constrained portfolio to allocate to it are different problems. AllocFlow targets the gap between them: it learns the portfolio as a joint action, so each position is chosen in relation to the rest of the book rather than assembled afterward from independent signals.
Key Findings
- A cross-asset flow Transformer generates signed weights for all eligible instruments together, modeling dependencies between positions.
- Training first imitates a strategy to establish structured candidate portfolios, then applies group-relative reinforcement learning to refine decisions using portfolio-level financial rewards.
- On hourly cryptocurrency perpetual futures, RL refinement raised reported net annualized return from 27.79% to 51.38% and improved Calmar from 0.955 to 4.043, while maximum drawdown fell from 29.10% to 12.81%.
- The policy reached a reported Calmar ratio of 4.90 on a further six months of 2026 data without retraining. These are results from the paper’s specified backtest and should not be read as evidence of live-trading performance.
How the Method Works
The system uses a frozen Kronos financial foundation model to encode recent OHLCV histories, but does not use its price forecasts as trading signals. A flow-based action head turns those representations into a joint set of target weights; a fixed-notional ledger then derives orders from changes in weights and accounts for trading fees. This separation makes the key idea clear: the learned policy optimizes the book and its financial reward, not just predictions about individual assets.
Who It Suits—and the Trade-off
A useful read for researchers comparing direct portfolio policies with forecast-then-allocate pipelines, especially those studying flow matching or reinforcement learning for finance. The results are specific to a historical cryptocurrency perpetual-futures evaluation, with its chosen universe, execution assumptions, and fees; readers seeking evidence across asset classes or live deployment should look elsewhere for validation.