Most price prediction topics start with a fixed list of assets. Someone picks BTC, ETH and SOL, a topic gets created for each, and every model is scored on that list whether or not it has anything useful to say about it.
The Hyperliquid Log Return topic, topic 90 on the Allora testnet, works the other way. It predicts three-minute log returns for Hyperliquid perpetual markets, and the set of markets being forecast is rebuilt every round from what the participating models actually submit. One topic, many markets, and a universe defined by where contributors choose to put a prediction.
What the target is
Each model submits a continuous prediction per market. The target is the natural-log return over a fixed interval.
r = ln(P_target / P_base) = ln(P_target) − ln(P_base)P_base is the price at the start of the interval and P_target is the price three minutes later. A positive value means the price rose, a negative value means it fell, and zero means no change.
Some reference points. A move from $100 to $101 is a ratio of 1.01 and a log return of about 0.00995. A move from $10 to $10.10 has the same ratio and the same log return. A fall from $100 to $99 is a ratio of 0.99 and a log return of about −0.01005.
Two properties make log returns the right unit for a multi-market topic.
The first is scale invariance. A 1% move produces the same value whether the asset trades at $0.40 or $40,000, so predictions across very different markets live on one comparable scale. That matters when a single submission can span dozens of markets.
The second is additivity over time. Log returns over consecutive intervals sum, which makes them easier to reason about and aggregate than simple percentage returns.
One practical warning. A log return is not a percentage return. A 1% move gives a log return of about 0.00995, close to 0.01 but not the same, and the gap widens as moves get larger. If your pipeline produces percentage changes or base-10 log returns, convert to ln(P_target / P_base) before submitting. Getting the unit wrong produces predictions that are directionally right and off in magnitude, which the scoring function will punish every round.
How the dynamic universe works
Topic 90 has no fixed asset list. A submission is a set of labelled values, where each key identifies a Hyperliquid market and the value attached to it is the predicted log return. The output shape is called MULTI, and the task is continuous regression rather than up, down or neutral classification.
Each round, the chain collects every non-default key contributed by active models into that round’s registry. When the submission window closes, the registry is frozen and becomes the set of markets that can be scored for that round.
This has a few consequences worth spelling out.
The universe can change round to round. If models start submitting predictions for a newly listed perp, that market enters the registry without anyone creating a new topic. If they stop, it drops out.
Multiple assets share one topic. Predictions for many markets are evaluated together rather than fragmented across a topic per asset, which keeps contributor attention and network weight concentrated instead of spread thin.
Key discovery is separate from price data. A key appearing in the registry defines what can be evaluated. Every evaluated key still needs its own ground truth.
The 100-key limit applies per submission. It caps how many markets one model can predict in a single round, while the topic as a whole can cover as many markets as contributors collectively submit.
There is also a quieter property here. In a fixed-universe topic, the asset list encodes the topic creator’s view of what matters. In the Hyperliquid Log Return topic, the registry encodes the contributors’ collective view of where a prediction is worth submitting. Whether that turns out to be a useful signal in its own right is one of the more interesting things to watch on testnet.
What “three minutes” means
The base time is the round’s starting block timestamp, also called the nonce time, rounded down to the full minute. The target time is exactly three minutes after that base. The return is always computed over a 180-second interval.
Take a round that starts at 09:41:33. The base is 09:41:00 and the target is 09:44:00. At the moment the round opens, 147 seconds remain until the target, and less remains as each submission arrives.
So the three minutes describes the base-to-target interval, and it says nothing about how much lead time a model gets after submitting or how far apart rounds are. Effective forecast horizon varies from round to round depending on where the round starts within its minute and when a model submits within the window. Part of the base interval has usually elapsed by the time you predict, which means some of the move you are forecasting is already observable. Models that account for that will likely behave differently from models that treat every round as a clean three-minute forecast.
The research forum has a longer discussion of the alignment choice, including why rounding to the minute matters for candle-based ground truth.
Ground truth and evaluation
Ground truth comes from Hyperliquid one-minute candle data, available through Atlas. For each key in the round’s registry, the reputer fetches the base and target prices at the defined timestamps and computes the log return. Predictions are then scored against that value, as described below.
For the comparison to be meaningful, predictions and ground truth have to agree on three things, the same price definition, the same timestamps and the same log base. If your training labels are built from a different price series, a different candle boundary or a different log base, your model is being trained on a slightly different target from the one it is scored on.
How a round is scored
Every market in the round is scored on its own. For each one, your predicted log return is compared with the realized log return using CZAR (Composite Zero-Agnostic Return), the topic’s loss function.
CZAR penalizes two mistakes most heavily, calling the wrong direction and undershooting the move. Overshooting costs less. This matters at a three-minute horizon. Realized returns sit very close to zero, so under a standard squared-error loss, a model that always predicts zero would be hard to beat. CZAR is designed so that a model with real directional skill doesn’t lose to a flat zero forecast. You can find the full explainer here.
Scoring covers the roughly 90 most liquid Hyperliquid perps by recent volume. Your score for the round is the average of your losses across those markets, with every included market weighted equally. The average is a geometric mean, taken in log space. That means one market where your loss blows up has less effect on the round than it would in a plain average.
Every model is averaged over the same set of markets. If a market is scored in the round and you didn’t submit a value for it, it’s filled with zero and scored as a no-change prediction. Skipping a market doesn’t take it out of your average. Under CZAR, a zero forecast on a market that moved counts as an undershoot, so leaving a market out is a bet that it stays flat.
In practice, that means three things. Submit for every market you have a view on. Make your forecasts commit to a direction. And don’t shrink your predictions toward zero to play it safe.
What it takes to be promoted
Promotion from testnet to mainnet is assessed per model and per topic. The Hyperliquid Log Return topic uses the same seven criteria as the other log-return topics. Predictions and ground truth are compared in log-return space, and the zero baseline, which represents no change, is what the WRMSE and WCZAR improvements are measured against.
Effective sample size. At least 20 effective observations.
Directional accuracy. The lower bound of the one-sided 95% confidence interval is above 50%.
Pearson correlation. The lower bound of the two-sided 95% confidence interval is above 0.
WRMSE improvement. The lower confidence bound is above 0%. WRMSE is root mean squared error weighted by the size of the realized move, measured against the zero baseline.
WCZAR improvement. The lower confidence bound is above 0%. WCZAR is the CZAR loss with the same weighting, measured against the same baseline.
Log aspect ratio. The confidence interval overlaps the range from −0.5 to +0.5, meaning the spread of your forecasts is not clearly too small or too large. Shrinking predictions toward zero to look safe fails this check, and so does inflating them to look decisive.
Participation. Strictly greater than 90%.
Most of these checks use the lower confidence bound rather than the headline value, because models have different numbers of submissions and a strong number on a short history is not yet evidence. On the Forge dashboard, a directional accuracy shown as 72.3% (44.7%, 100%) has a lower bound of 44.7%. That does not clear the 50% bar, however good the headline looks.
Full criteria are in the Forge docs, and the promotion and relegation policy is on the research forum. You can find the criteria here and the policy here.
Submission rules that change model behavior
These are the rules most likely to catch people out.
A submission can contain up to 100 market keys. A submission with more than 100 keys is rejected in full, so trim before you send.
Negative values are allowed, as you would expect for returns.
A value of exactly zero is dropped from the submitted key set. Zero is a valid no-change return mathematically, but the protocol treats it as the default value, so it does not count as an explicitly contributed key. If you are the only model predicting a market and you submit zero, that market will not enter the registry through your submission.
Missing values are filled with zero on chain. If a key is in the registry because another model submitted it, and your submission does not include it, your prediction for that key is recorded as zero. The reputer supplies ground truth for those keys but does not perform the fill.
Taken together, zero is the default prediction for every market in the registry. Staying silent on a market is equivalent to predicting no change, which is the random-walk baseline. That makes the baseline explicit. Because scoring averages over every scored market in the registry, the markets you skip still count, as zero forecasts.
Block timing
The topic runs with an epoch length of 36 blocks, a ground truth lag of 36 blocks and a worker submission window of 6 blocks.
These are block counts, so their wall-clock durations vary with block production. The submission window controls when models can send predictions. The ground truth lag controls when reputation reporting becomes available, and it includes a buffer so evaluation does not arrive before the target candle exists. Neither parameter changes the 180-second interval used to compute the target return.
What is hard about this topic
Three-minute returns on perpetual markets are mostly noise. The expected signal-to-noise ratio at this horizon is low, and most individual models should expect to beat the zero baseline by small margins if at all. That is why promotion is judged on confidence bounds rather than headline numbers, and it is also the reason the topic is interesting as a network problem. Small, weakly correlated edges across many models and many markets are the setting where aggregation can add value that no single model delivers, and it is also the setting where overfitting to a backtest is easiest.
A few open questions we expect testnet to help answer.
How does the registry evolve? Do models concentrate on the most liquid markets, spread across the long tail, or specialise?
How do thinner markets inside the scored set behave? One-minute candles on illiquid perps are noisier, and a market with few predictors has less to aggregate.
How much does variable lead time matter? Models that condition on the elapsed portion of the interval may separate from those that ignore it.
Does a contributor-defined universe carry information? Which markets models choose to predict, and when, may itself say something about where short-horizon edge is available.
Build against the Hyperliquid Log Return topic
The Hyperliquid Log Return topic is live on the Allora testnet as topic 90. If you already run short-horizon return models, the main adjustments are the natural-log target, the minute-aligned base time, the 100-key cap, the zero-default behavior and the CZAR loss. Deploy a model against it on Forge and see how it scores against the rest of the network.
