The Allora Forge platform has introduced a new prediction category, designated as topic 90, focused on Hyperliquid perpetual futures. The system went live on October 8, 2026, and tasks machine learning models with forecasting three-minute log returns. Unlike static setups, the list of markets is dynamic and rebuilt for every round.
Models can submit forecasts for up to 100 markets per round. The platform uses the Composite Zero-Agnostic Return (CZAR) loss function to evaluate performance, which rewards models for taking clear directional stances while penalizing those that avoid making specific predictions. Ground truth data is sourced from Hyperliquid’s one-minute candle data via the Atlas platform.
To qualify for mainnet deployment, a model must pass seven distinct statistical tests. One requirement is that the lower confidence bound on directional accuracy must exceed 50 percent, ensuring the model performs better than random chance. These measures are intended to verify the credibility of predictive systems.
Currently, more than 2,000 models are active on the Allora Forge platform, covering 18 different topics including price, volatility, and commodity forecasts. The new integration with Hyperliquid does not involve the issuance of new tokens or changes to existing trading mechanisms, focusing instead on the performance of predictive algorithms.
