Nov 12 2025: Major Changes in Numerai Crypto Tournament

Numerai has released the Crypto V2 “Spectra” Dataset, bringing a major overhaul to the crypto tournament workflow. This update introduces a smaller, more curated universe of 300 coins and provides built-in technical analysis features, including moving averages, RSI, momentum, volatility, and Bollinger bands with 20- and 60-day variants.

The targets in historic data have also been updated: participants now have ranked & binned 20-day and 60-day return targets. The live universe does not includes these targets, they exist only in historical training data. Participants must produce a single prediction per symbol for the live universe, which Numerai then scores against both horizons.

The update fundamentally changes the way the contest is approached. Unlike the previous V1 setup, users no longer need to source hundreds of altcoins, compute technical indicators, or filter coins by correlation. The workflow is now different: data preparation now focuses on fusing your own signals (for example, from Yiedl) with Numerai’s provided features, aligning to the official universe, and producing a fully enriched dataset for modelling. The modelling pipeline itself can remain largely unchanged, but the pre-processing and feature alignment steps must now follow this new structure to produce valid submissions.

In short, the Crypto contest has become less DIY and more standardized.

---

What is Yiedl data?

Numerai crypto contest is an excellent opportunity for anyone interested in ML, AI trading bots, and finance. If you enjoy mathematics, trading or predictive modeling, it's worth checking out! That said, the learning curve is steep due to the lack of structured documentation and the enigmatic nature of their targets (target is black-box but related to 30-day returns). To remind you, we are trying to predicting the Numerai target but we do not know what it really is. Why do that at all? Because Numerai allows you to benchmark your predictive modelling! It is essentially a live testing system for your financial forecast models, free of charge. Treat it like this instead of a shortcut to profit, and your mental wellbeing will be just fine.

So, my story. I had been doing some financial modelling before, only as an academic interest in mathematical modelling. However, I was always sceptical about back-testing due to overfitting conciderations. I wanted just that - an easy, systematic way to test my predictions in a live environment. I though, haha, I can piggyback Numerai for that. Yes, fellows, it is possible, but you will have to play the game by their rules -- meaning: predicting THEIR target, not yours.

So my experience with Numerai crypto, unlike I originally expected, started with a lot of frustration: finding data, making sense of the targets, figuring out the best submission workflow was not easy. In the crypto contest, you need to source your own data, and the available historic target data is sporadic (yes, values are not for every single day of the calendar, there are huge gaps for some symbols), which makes the challenge even more difficult. And you actually need to have a very decent DevOps qualification just to run all of the shit every day of the year. If you’re serious about participating, you must decide which data sources to use and you will have to develop custom scripts to process data efficiently.

Initially, for building my own model, I attempted to collect price data using various price APIs (also from Binance, among other sources, where I understood that Binance data from US and EU are different datasets!!!), running a cron job on a VPS to accumulate historical data, store it on my server, and train models on it. Numerai requires at least 100 tradable assets per submission daily. So I quickly figured out that 100 signals per submission weren’t enough, likely due to their strict requirements on non-correlated assets. In practice, a single valid submission typically needs at least 200-300 symbols, meaning daily predictions for that many crypto assets. Again: yes, you must submit at least 200-300 ticker predictions for a single model.

So this is where Yiedl data is useful. It easily meets this requirement, covering hundreds of assets out of the box, and it's a single point of accessing the data. I highly recommend looking into Yiedl data. Numerai partnered with Yiedl.ai to provide high-quality crypto datasets with over ten years of historical observations. While the dataset is obfuscated to protect IP, it's a valuable resource for model training. If you're benchmarking financial prediction models on Numerai, it is crucial to know about the community data provided by YIEDL.ai.

Numerai partnered with YIEDL to bring a comprehensive crypto dataset to Numerai Crypto. This dataset includes over ten years of cryptocurrency data, covering essential features like price volume momentum (PVM), sentiment analysis, and on-chain metrics. Just like the obfuscated data used in the classic Numerai tournament, the YIEDL dataset is encrypted to protect intellectual property while remaining highly useful for model training.

YIEDL provides two versions of its dataset:

• Historical Dataset: Contains records dating back to 2013, making it invaluable for training models using long-term crypto trends

• Latest Dataset: Covers the most recent month, optimized for quick access

While YIEDL datasets are very high-quality and usable, keep in mind that different symbols have varying time spans of coverage. Additionally, due to its massive size, managing the dataset can be challenging, requiring custom scripts for effective processing of parquet and csv files mixed together. However, even despite the obfuscation, the datasets are a valuable asset for building robust prediction models on Numerai. My workflow is to have historic data regularly updated with the latest data files, so that in such way you only need to download the huge historic data file infrequently.

There is not much info and documentation on data science contests like Yiedl or Numerai , but in terms of profitability one can judge that the contests serve their purpose. For example, Numerai recently announced, that "Numerai Crypto is far more profitable for users than our other 2 tournaments" that they host.

To take part in the Numerai crypto contest one needs to get the data first. Where from? - It is really up to you: you will need to decide which data to use, how to use it, where to collect it, and so on. Contact me if you need a consultant on machine learning operations (mlops) or devops, data collection, parsing and preparation tasks or if interested in use cases for these particular scripts.

Related

Check out repo 🔗Intuitive Trader Game: gamification of prediction in finance.

What is included?

This repository contains unofficial utilities for the Numerai Crypto Signals Contest. Intended to help users interact with the Numerai API for various purposes related to the crypto signals competition only.

Requirements

To use these utilities, you need to install the numerapi package:

pip install numerapi

Scripts

round.py

This script fetches the current round information from the Numerai API. It provides details about the ongoing round, including start and end times, and other relevant information.

model.py

This script fetches information about a specific model by its title. It retrieves various details, including the model ID, latest submission information, stakes, and other relevant metrics.

submit.py

This script automates the submission of crypto signals to https://crypto.numer.ai/submissions. It streamlines the process of submitting your signals, making it easier to participate in the Numerai Crypto Tournament.

scrap.py

Unfortunately, Numerai does not currently provide an official documented API to access batch Resolved Rounds data for crypto models. This script is a screen scraping tool that automates downloading and parsing this data from the web interface. We apologize for relying on scraping techniques, which are less reliable and more brittle than an official API. We hope Numerai will add a supported API endpoint for this data in the future to enable easier and more robust access.

Usage: python script.py model1,model2,model3"

This will save csv file with resolved rounds statistics for your models.

Yiedl Extractor

yiedl-extractor.py is a script for extracting specific columns from the historical Yiedl dataset. This extracted data is used for the Numerai data contest.

Features

• Loads historical dataset from a Parquet file.
• Extracts user-specified columns.
• Filters data starting from a defined date.
• Saves the filtered dataset as a CSV file.

Usage

1. Download the historical dataset from:

https://api.yiedl.ai/yiedl/v1/downloadDataset?type=historical

(Data is released every Monday to Friday after 12 UTC.)
2. Update `PARQUET_FILE` with the correct dataset filename.
3. Modify `COLUMNS_TO_KEEP` to specify the columns you want to extract.
4. Set `START_DATE` to filter data from the desired start date.
5. Run the script:
```sh
python yiedl-extractor.py

Configuration

• Input Parquet File: dataset_historical_20250201.parquet (Update to the latest dataset file.)
• Output CSV File: yiedl_data.csv
• Columns to Extract: ['date', 'symbol', 'pvm_XXXX'] (Modify as needed.)
• Start Date: 2023-01-01 (Adjust to filter relevant data.)

Output

• The extracted data is saved in yiedl_data.csv.

Notes

• Ensure the dataset file exists in the specified directory before r