Allora’s Forge platform has opened a new prediction topic built around Hyperliquid, and it comes with an unusual twist: there is no fixed list of assets.
The Hyperliquid Log Return topic, designated topic 90, went live on October 8, 2026. It asks AI and machine learning models to forecast three-minute log returns across Hyperliquid’s perpetual futures markets, with the set of markets rebuilt every round.
How topic 90 actually works
Topic 90 works differently from static prediction setups. The market set for each round is assembled dynamically, based on what models choose to submit.
Each submission is capped at 100 keys. A key, in this context, is essentially a market the model wants to weigh in on.
Zeros are treated as the default and dropped from the registry. So a model that has no opinion on a given market simply leaves it out, rather than padding its entry with empty guesses.
The target itself is a three-minute base-10 log return. A log return measures how much a price moved over a window, expressed as a logarithm of the ending price divided by the starting price.
To settle who was right, Allora pulls ground truth from Hyperliquid’s one-minute candle data through its Atlas platform.
The scoring system punishes fence-sitting
Allora’s answer to zero-padding is the CZAR loss function, short for Composite Zero-Agnostic Return. It is designed to reward forecasts that actually take a directional stance and to penalize predictions that hedge toward nothing.
Models are assessed on several criteria, including directional accuracy and correlation with realized returns.
The bar for graduating is also explicit. To move to mainnet, a model must pass seven statistical tests per topic and worker.
One of those tests requires the lower confidence bound on directional accuracy to sit above 50%. That means a model has to show, with statistical confidence, that it beats a coin flip at calling direction.
Forge, Hyperliquid and the bigger picture
Forge functions as a competitive arena where models build reputations through testing. According to the research findings, more than 2,000 models are already active on the platform.
Those models are spread across 18 topics. The Hyperliquid Log Return topic joins others focused on price, volatility and commodity forecasts.
Hyperliquid is a high-performance Layer 1 blockchain optimized for perpetual futures and spot trading. Its HyperCore infrastructure handles order matching and liquidations.
Notably, the launch does not introduce a new token or any new trading mechanism. It is a model-performance play: a new arena with new rules, aimed at the people building predictive systems rather than the people trading on Hyperliquid directly.
What this means for model builders and traders
The dynamic market set is the most interesting design choice here. By letting models pick their own markets, up to 100 at a time, Forge shifts some of the strategy from “how good is your forecast” to “where are you confident enough to forecast at all.”
The CZAR loss function and the seven-test mainnet gate address a credibility problem common in AI prediction projects. Requiring a directional accuracy lower bound above 50% forces models to demonstrate an edge, not just low error.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.