Hyperliquid Ranking

Zero to Submission in 20 Seconds

Go from setup to your first prediction submission with our end-to-end tutorial notebook.

Objective

The Hyperliquid Ranking Challenge requires participants to rank crypto assets on the Hyperliquid decentralized derivatives exchange by their expected relative returns over the next 10 and 30 days. The challenge universe comprises approximately 165-175 (and likely more in the future) liquid tokens on the Hyperliquid protocol. This universe may change periodically, with tokens added or removed to ensure it remains as actionable as possible. If a token does not have enough volume or liquidity, it will likely be removed from the universe.

from crowdcent_challenge import ChallengeClient
client = ChallengeClient(challenge_slug="hyperliquid-ranking")
client.get_challenge() # Get more challenge details

Training data

The training dataset is created just to get you started. Simple models can be built with just the features and targets, but don't expect to win the challenge with them. We recommend building your own training datasets with sources like ccxt, eodhd, coingecko, or yfinance.

You can download our training data, including features and targets from crowdcent.com/challenge/hyperliquid-ranking or via the CrowdCent client:

client.download_training_dataset("latest")

id	eodhd_id	date	feature_1_lag15	feature_1_lag10	...	feature_1_lag0	target_10d	target_30d
BABY	BABY32198-USD.CC	2024-03-20	0.123	0.145	...	0.823	0.15	0.25
OM	OM-USD.CC	2024-03-20	0.456	0.423	...	0.756	0.35	0.45
IOTA	IOTA-USD.CC	2024-03-20	0.789	0.812	...	0.923	0.55	0.65
MOODENG	MOODENG33093-USD.CC	2024-03-20	0.234	0.267	...	0.445	0.75	0.85
ENS	ENS-USD.CC	2024-03-20	0.567	0.534	...	0.678	0.95	0.88

Asset IDs

We currently provide id (the hyperliquid id) and eodhd_id (the id to download via EODHD) for each asset. You can request we include additional id mappings from other data vendors in the inference data if data licenses allow.

If you're using CrowdCent's training data to build your models, the inference data features will always match that of the latest training dataset version. The assets are generally the same as you would find in the training data, but new additions or removals are possible. For the inference data, we aim to track the listed and tradeable perps on Hyperliquid.

Features

The training data contains 80 total features following the pattern feature_{n}_lag{lag}:

20 unique features (n = 1 to 20)
4 lag values per feature (0, 5, 10, 15 days)

Features with the same number represent the same metric at different time points. For example, feature_1_lag0 through feature_1_lag15 track the same underlying metric over time. This structure preserves temporal relationships, allowing you to build sequence models (LSTM, GRU, Transformer), identify trends/patterns across different time horizons, and engineer additional features based on temporal changes.

Targets

Targets are the rankings of an asset's 10d and 30d forward relative returns (with a 1d lag). Targets do not currently take funding rate or any other factors (e.g. market cap, volume, etc.) into account. It's possible that the targets will be updated in the future to include such factors.

Why the 1-day lag? For our purposes, the crypto universe has a close time of 24:00 UTC. At 14:00 UTC when predictions are made, only data through the previous day's 24:00 UTC close is available. The lag ensures predictions use only historical data while forecasting returns starting from tonight's close.

Timeline:

Day D-1: 24:00 UTC → Close price finalized (latest available data)
Day D:   14:00 UTC → Inference pipeline starts
         14:00-18:00 UTC → Inference period lasts 4 hours
         24:00 UTC → Prediction/Scoring period starts (Day D close)
Day D+10/30: 24:00 UTC → Prediction/Scoring period ends

Predictions rank assets by expected performance over the next 10/30 days starting from tonight's close
Rankings are relative. They say nothing about the expected absolute performance of an asset.

Inference data

Inference Period Open: The internal pipeline starts at 14:00 UTC. The file usually becomes available a few seconds to a few minutes later, once data quality checks pass.
Inference Period Close: 4 hours after the actual release timestamp (typically around 18:00 UTC).

Each day, an inference dataset is released containing the universe of tokens for which predictions are required. The inference data contains features but has no targets as they do not exist at the time of your submission. Your predictions will be scored against resolving targets from real market data in the future.

Polling and parameter options

Current vs Latest: Use "current" to download today's active inference period (for making submissions). Use "latest" to download the most recently published period (which may be from a previous day if today's isn't ready yet).

Polling: The inference file may not exist the very instant the clock strikes 14:00 UTC. If you use release_date="current" with default parameters, polling is handled automatically. For release_date="latest", no polling is needed since it always fetches the most recent available period.

For manual polling, use client.wait_for_inference_data("inference_data.parquet") or gently poll the API until GET /current stops returning 404.

To download inference data:

# Download the current active inference period (for today's submissions)
client.download_inference_data("current") # THIS WILL FAIL IF THERE IS NOT AN OPEN INFERENCE PERIOD

# Download the most recently available inference period
client.download_inference_data("latest")

# Download a specific date's inference data
client.download_inference_data("2024-12-15")

id	eodhd_id	feature_1_lag15	feature_1_lag10	feature_1_lag5	feature_1_lag0	...
BABY	BABY32198-USD.CC	0.123	0.145	0.112	0.156	...
OM	OM-USD.CC	0.456	0.423	0.467	0.401	...
IOTA	IOTA-USD.CC	0.789	0.812	0.798	0.823	...
MOODENG	MOODENG33093-USD.CC	0.234	0.267	0.223	0.289	...
ENS	ENS-USD.CC	0.567	0.534	0.578	0.512	...

Tip

You do not need to use the features included in the inference data. You can use and are encouraged to use your own features. It may still be helpful to use our inference data and id mappings to use the same universe of assets.

Submitting predictions

Minimum of 80 ids from the inference data are required for a valid submission. The following columns are also required (no index):

id: The id of the asset on Hyperliquid.
pred_10d: A float between 0 and 1 representing the predicted rank for the 10-day horizon.
pred_30d: A float between 0 and 1 representing the predicted rank for the 30-day horizon.

To submit predictions, you have 5 submission slots available. These can be used to submit multiple predictions for the same day and are defined by the slot parameter:

client.submit_predictions(df=predictions_df, slot=1) # Submit a dataframe
client.submit_predictions(file_path="submission.parquet", slot=2) # or a parquet file

Flexible Timing

You can submit anytime. If the window is open, your submission is accepted immediately. If the window is closed, your submission is queued and automatically submitted when the next window opens. By default, submissions are also queued for the following period (auto-rollover) if you submit during an open window. Use queue_next=False to opt out.

id	pred_10d	pred_30d
BABY	0.2	0.3
OM	0.4	0.5
IOTA	0.1	0.2
MOODENG	1.0	1.0
ENS	0.3	0.4

Note

If you are submitting a dataframe, id must be a column in the dataframe, not the index. If you are submitting a file, all submissions must be in parquet format.

Scoring and Evaluation

Before scoring, for each prediction timeframe, ids are uniform ranked [0, 1], and any missing ids are filled with 0.5.

Raw Metrics

These metrics measure how accurate your predictions are against actual market outcomes.

1) Symmetric Normalized Discounted Cumulative Gain (NDCG@40)

When you see NDCG@40, think: "how well did I rank the top 40 assets and how well did I rank the bottom 40 assets?" With ~170 tokens in the universe, k=40 represents approximately the top/bottom 20-25% of assets. This metric equally rewards both:

Top 40 identification: Finding the tokens that will have the highest returns (for long positions)
Bottom 40 identification: Finding the tokens that will have the lowest returns (for short positions or avoidance)

The logarithmic discount means getting the #1 ranked token correct is much more valuable than getting the #40 ranked token correct. A perfect NDCG@40 score of 1.0 means you perfectly ranked both tails of the distribution. This metric is particularly valuable for portfolio construction where you want to maximize exposure to the best performers while avoiding or shorting the worst.

Random Baseline

Random predictions score approximately 0.55 for NDCG@40 with ~170 tokens, not 0.5. See the detailed explanation for why this happens.

2) Spearman Correlation

Spearman's rank correlation (ρ) measures how well your predicted ranks align with the true ranks across the entire universe of ~170 tokens. Unlike NDCG@40 which focuses on the 40 extremes, ρ treats all rank positions in the entire universe equally.

Uniqueness Metrics

In addition to raw accuracy, CrowdCent evaluates how unique your predictions are relative to the meta-model. The meta-model already captures the crowd's collective wisdom -- uniqueness metrics measure whether you're contributing new predictive signal that the crowd doesn't already have.

The leaderboard provides a Raw / Unique toggle to switch between these two views.

Three uniqueness metrics are computed for each horizon (10d, 30d):

1) Unique Spearman (unique_spearman) -- Your predictions are first neutralized against the meta-model to isolate the orthogonal component, then scored with Spearman correlation against actuals. Positive values mean your unique signal is predictive.

2) Unique NDCG@40 (unique_ndcg@40) -- Same neutralization step, but scored with Symmetric NDCG@40. Measures whether your unique signal correctly identifies the top and bottom 40 assets.

3) Correlation to Meta (corr_to_meta) -- Spearman correlation between your predictions and the meta-model. Measures how similar your signal is to the crowd. Values closer to zero indicate more differentiated predictions. Unlike the above two, this is a pure similarity measure and is not included in the unique composite percentile.

For detailed explanations of the math behind these metrics, see Uniqueness Metrics.

Why Uniqueness Matters

A submission that's highly correlated with the meta-model adds little new information, even if its raw scores are strong. Conversely, a submission with moderate raw scores but high uniqueness can significantly improve the meta-model by contributing orthogonal signal. The best participants deliver both accuracy and differentiation.

Composite Percentile

Since NDCG@40 (0-1 range) and Spearman correlation (-1 to 1 range) have different scales and distributions, we use a composite percentile for fair comparison.

The composite percentile is calculated as the average of your percentile rankings across the four raw metrics:

\[ \text{Composite Percentile} = \frac{1}{4} \times \left( \begin{array}{l} \text{percentile(NDCG@40}_{10d}) + \\ \text{percentile(NDCG@40}_{30d}) + \\ \text{percentile(spearman}_{10d}) + \\ \text{percentile(spearman}_{30d}) \end{array} \right) \]

Where each percentile represents your ranking (0-100) compared to other participants for that specific metric for a given day/inference period.

Note

The composite percentile uses raw metrics only. Uniqueness metrics have their own composite — see Unique Composite Percentile below.

Important: Composite percentiles are only calculated when ten (10) or more valid submissions (counted by submission slots, not users) are received for a given day. If fewer than ten submissions are present, the composite percentile will not be calculated, and you'll need to look at absolute metric scores instead.

As we learn more about the challenge's metamodel, we may adjust the weighting or add/remove metrics.

Unique Composite Percentile

Analogous to the composite percentile, but for uniqueness metrics. The unique composite percentile averages your percentile rankings across the four unique metrics:

\[ \text{Unique Composite Percentile} = \frac{1}{4} \times \left( \begin{array}{l} \text{percentile(unique\_NDCG@40}_{10d}) + \\ \text{percentile(unique\_NDCG@40}_{30d}) + \\ \text{percentile(unique\_spearman}_{10d}) + \\ \text{percentile(unique\_spearman}_{30d}) \end{array} \right) \]

This gives you a single number capturing how well your differentiated signal performs relative to other participants. It is visible in the Unique leaderboard view and profile sparklines.

The unique composite percentile is kept separate from the raw composite percentile so that the raw composite can always serve as a fair comparison against the meta-model. The unique composite percentile does not currently factor into CC Points calculations.

Score Ranges and Percentile Rankings

Raw Score Ranges:

NDCG@40: Ranges from 0.0 (worst possible) to 1.0 (perfect ranking of top and bottom 40 assets)
ρ (Spearman's Rank Correlation): Ranges from -1.0 (perfect inverse ranking) to 1.0 (perfect ranking), with 0.0 indicating random performance

Uniqueness Score Ranges:

unique_ndcg@40: 0.0 to 1.0 (~0.55 = random baseline, higher = better)
unique_spearman: -1.0 to +1.0 (higher = better unique signal)

Why Are Scores Typically Low?

Financial markets are characterized by extremely high noise-to-signal ratios. Seemingly "low" scores can be quite competitive in this domain. Additionally, market regimes shift over time, causing the distribution of achievable scores to fluctuate significantly.

Percentile Rankings: Your Most Reliable Metric

CrowdCent calculates percentile rankings that show where you stand relative to other participants. These percentiles are recalculated daily for both raw and uniqueness metrics.

Tracking your percentile scores over time is often more informative than focusing on absolute scores, as it accounts for evolving competition and regime shifts that affect all participants. A model that consistently ranks in the 75th percentile across different market conditions can often be more valuable than one that occasionally achieves top scores but performs poorly in other regimes.

Minimum submissions for percentile

Percentiles only calculated when ten (10) or more valid submissions are received for a given day. If fewer than ten submissions are present, you'll see individual metric scores, but no percentile.

Meta-Model

The CrowdCent Meta-Model aggregates predictions from all participants, representing a "wisdom of the crowd." It serves two purposes:

Downloadable signal: Real-time predictions available to Challenger+ tier (100+ CC Points). All users can access 90-day delayed predictions and full performance scores.
Uniqueness benchmark: Used as the reference for computing uniqueness metrics -- your predictions are compared against it to measure differentiation.

This may change in the future with no notice.

Meta-Model Disclaimer

Meta-model signals are released for informational and educational purposes only. Not financial, investment, or trading advice. CrowdCent disclaims all liability for any losses, damages, or consequences arising from use of the meta-model. Users assume all risks.

Construction Methodology

The meta-model is constructed daily using a points-weighted average of all submissions:

Uniform Ranking: Each individual submission's predictions are first converted to uniform rankings [0, 1] for each prediction column (pred_10d, pred_30d)
Missing ID Handling: Any asset IDs missing from individual submissions are filled with neutral rankings of 0.5 after the uniform ranking step
Average slots: Create a single prediction for each user by taking the arithmetic mean of all normalized rankings across all submission slots for each user.
Points-Weighted Average: The final meta-model is created by taking a weighted average across all users, where each user's weight is proportional to their CC Points EMA (Exponential Moving Average with 7-day half-life).

Users with more accumulated CC Points have greater influence on the meta-model. See CC Points System for details on how points are earned and how weights are calculated.

Access and Downloads

Real-time meta-model predictions require Challenger tier or above (100+ CC Points). Users below Challenger can access predictions on a 90-day delay. Performance scores are always available to everyone.

The meta-model is available through multiple channels:

Via web: https://crowdcent.com/challenge/hyperliquid-ranking/meta-model/
Via API: client.download_meta_model(dest_path="meta_model.parquet")

The meta-model is a parquet file with the following columns. New predictions are added daily, creating a time series with multiple release dates as shown in this sample:

id	pred_10d	pred_30d	release_date
BTC	0.85	0.82	2024-12-15
ETH	0.74	0.78	2024-12-15
SOL	0.91	0.89	2024-12-15
BTC	0.85	0.82	2024-12-16
ETH	0.74	0.78	2024-12-16
SOL	0.91	0.83	2024-12-16
...	...	...	...