Numerai is a blockchain-based crowdsourcing AI quantitative hedge fund platform. Data scientists participate in the global stock market prediction competition by training machine learning models to predict encrypted financial data. The contributed models are aggregated into a meta model (Meta Model) through the NMR (Numeraire) token mortgage mechanism, which is used to drive the trading decisions of real hedge funds. The platform was founded by Richard Craib in 2015 and has received investments from Union Square Ventures, Paul Tudor Jones, Renaissance Technologies co-founder Howard Morgan, etc. It will complete $30 million in Series C financing in 2025, with a valuation of $500 million.

numerai

Core parameters and statistics

Numerai is a unique platform that combines data science competitions with quantitative hedge funds. It provides high-quality financial data sets that have been encrypted and obfuscated to data scientists around the world for free. Participants train machine learning models to predict encrypted stock return indicators. The submitted predictions are aggregated into a meta model (Meta Model) through the NMR (Numeraire) token mortgage mechanism, which is ultimately used by the Numerai hedge fund for real market transactions.

Projects	Public Information
Official Positioning	Crowdsourcing AI Quantitative Hedge Fund and Data Science Competition Platform
Founder and date of establishment	Richard Craib, October 2015
Headquarters	San Francisco, USA
Core Mechanism	Global Data Scientist Competition → Meta-Model Aggregation → Hedge Fund Trading
Latest data set	V5.2 Faith2 (2025-12, 186 new features)
Total NMR payouts	$24,400,000+ (data scientists earned cumulatively)
Number of staked models	4,500+
Fund management scale	~$550M (2025-11), net return in 2024 25.45%
Investors	Union Square Ventures, Paul Tudor Jones, Howard Morgan (Renaissance Technologies)
GitHub Stars/Forks	1,200+ / 310 (example-scripts repository)
Available platforms	Web, API, Python/numerapi libraries

Data Features: The data provided by Numerai has undergone "structure-preserving encryption". The features are anonymized but statistical correlation is retained. Data scientists can model without a financial background. Each piece of data corresponds to a certain stock at a specific time point (era), characterized by an encrypted value of a quantitative attribute, and the target is an abstract return measure for the next 20 trading days.

Metamodel aggregation logic: Only NMR-collateralized predictions will be included in the metamodel. The staking mechanism ensures that participants have an economic incentive to submit high-quality predictions - NMR is rewarded for good performance, and NMR is burned for poor performance. This "skin in the game" design allows hedge funds to trust predictive signals from anonymous third parties.

User and market recognition

Numerai's market recognition is reflected in three dimensions: institutional investors, top financial institutions, and the global data scientist community.

Institutional investment: Completed $30 million in Series C financing in November 2025, with a valuation of $500 million, a five-fold increase from the 2023 valuation. It was led by top university endowment funds, with participation from Union Square Ventures, Shine Capital, Paul Tudor Jones, etc. In August 2025, JPMorgan Asset Management invested money in Numerai and acquired $500 million in capacity, which was one of the largest bets on AI-driven quantitative strategies by a traditional top financial institution.

Fund Performance: 2024 was Numerai's best year ever, with the Global Equity Market Neutral hedge fund posting a 25.45% net return, a Sharpe ratio of 2.75, and only one month in the red for the year. AUM has grown from approximately $60 million to $550 million over the past three years. Against the backdrop of a general cryptocurrency market decline of over 90% in 2022, Numerai still achieved a positive return of 20% and returned earnings to investors.

Data Scientist Community: The platform has 4,500+ staked models and has paid out more than $24.4 million to data scientists. The example-scripts repository on GitHub has 1,200+ stars and 310 forks, the Discord community is active, and NumerCon is held every year.

Cost advantage

Numerai’s cost structure presents completely different characteristics at three levels: C-side data scientists, API developers, and institutional investors.

C-side/Data Scientist: Entry is completely free. Anyone can register for an account and download the encrypted dataset for free, train a model using Python or R and submit predictions. Participate in leaderboards and earn performance ratings without staking NMR. Only when participants choose to stake NMR (requiring self-purchased tokens) to obtain economic incentives do they face real funding risks - poor model performance will cause NMR to be burned. For users who just want to experience and learn, the zero-cost threshold is extremely low.

API/Developer: Numerai provides a Python library (numerapi) and REST API for automated data downloads, prediction uploads, and model management. The official also provides an MCP (Model Context Protocol) interface, which allows AI Agents to interact directly with the platform - such as using natural language instructions to complete the entire process from modeling to submission. These API interfaces are free to use, with no limit on the number of calls and are only subject to the fair use policy.

Institutional Investors: The fund level is open to accredited investors. The investment threshold and fee structure are not disclosed and need to be confirmed by contacting investing@numer.ai. The $500 million in capacity that JPMorgan received in 2025 implies strong demand for funds at the institutional level, but for individual investors, the threshold for direct investment funds is higher.

Hidden Costs: The main cost faced by participants is time - data science competitions require a lot of effort in feature engineering, model tuning, and continuous iteration. In addition, fluctuations in the market capitalization of the NMR token itself can amplify the uncertainty of actual gains/losses.

Main functions

• Data Science Competition Platform: Core function, providing encrypted financial data sets (three sets of training/validation/real-time), data scientists train models to predict target variables, and submit prediction files to participate in the leaderboard rankings. The competition releases new live data every Tuesday through Saturday.

• NMR Staking and Incentive Mechanism: Participants can stake NMR tokens on their own models. Each round of staking lasts for approximately 20 trading days (i.e., one scoring cycle). Models with positive performance receive NMR rewards, while models with negative performance have part of their NMR burned. This design ensures that the metamodel aggregates only high-quality predictions.

• Meta Model: Numerai's core asset. All pledged predictions are combined into a meta-model by weight and fed directly into the hedge fund's trading system. The meta-model weight is dynamically adjusted with the Scores (CORR and MMC) of each model, similar to the weight optimization process of neural networks.

• Numerai Signals: An independent competition platform that allows data scientists to build stock prediction signals using their own private data (rather than the encrypted data provided by Numerai). Suitable for participants who already have their own data sources and professional financial knowledge.

• Crypto Tournament: an independent competition for the cryptocurrency market, designed more for entertainment and education, not for actual trading.

• AI Scientist / MCP Agent Integration: A new capability launched in 2025-2026, data scientists can use the MCP protocol to allow AI Agents (such as Codex, Claude, Cursor) to directly participate in the modeling process - from data download, feature analysis, model construction to prediction submission, all completed through natural language instructions.

• Numerai Bounties: A community bounty mechanism, Numerai releases specific research directions or modeling challenges, and participants can receive NMR rewards for completing them.

Model and version evolution

Numerai's "version" system is mainly reflected in the iteration of data set versions and the milestone evolution of platform functions.

Dataset version

The dataset is the core asset of the Numerai competition, evolving iteratively in the V series:

• V5.2 Faith2 (2025-12): The latest version, which introduces 186 new features and adds two new target variables: Ender and Jasper (each containing two windows of 20 days and 60 days). All data uses int8 format to compress storage and speed up loading.
• V5.1 (~2025-06): Intermediate incremental update, feature set fine-tuning and optimization.
• V5.0 (~2025-01): The first version of the V5 series, the feature architecture has been significantly updated, laying the foundation for subsequent expansion.
• V4.2 (~2024-06): The final version of the V4 series, with multiple feature and target adjustments.

After each data version is released, support for the old version will be gradually stopped, and contestants need to migrate to the latest version to remain competitive.

Platform feature milestones

• 2015-10: Numerai was officially established and received $1.5 million in seed round financing (FT report).
• 2016-12: The crowdfunding hedge fund concept is officially open to machine learning experts (TechCrunch reports).
• 2017-02: Published NMR token white paper to introduce blockchain into data science competition incentives (Wired report).
• 2022: AUM grows from zero to ~$60M, maintaining positive returns amid market headwinds.
• 2023: Financing round valued at approximately $100 million.
• 2024: Fund achieved 25.45% net return, Sharpe 2.75, rapid AUM growth.
• 2025-08: JPMorgan invests and secures $500M capacity, Landmark Institutional Recognition.
• 2025-11: Completed $30 million in Series C financing, valued at $500 million, and AUM reached $550 million.
• 2025-12: Release of V5.2 Faith2 dataset.
• 2026-02: Hosting NumerCon 2026 conference.
• 2026-07: Launched the third strategic NMR repurchase to promote the Atomic Blockchain Staking function.

Technical advantages

Numerai's technical architecture is designed around the core challenge of "how to aggregate predictive signals from untrusted third parties while protecting data privacy."

Structure-Preserving Encryption: Numerai applies encryption transformations to the original financial data so that the feature values ​​maintain statistical relevance (e.g. ordering, distribution shape) but cannot be restored to the original stock identification or specific financial indicators. This means that (1) the data can be freely distributed to anyone without fear of data leakage, (2) modelers do not need any knowledge in the financial field to participate, and (3) the model can only be used in Numerai competitions and cannot be "traded on one's own."

Meta Model weight optimization mechanism: Numerai does not simply take the average of all predictions, but operates through a mortgage-driven dynamic weight system. Each model's CORR (correlation coefficient with the target) and MMC (marginal contribution to the meta-model) determine its weight in the next round. The weights of poor-performing models are reduced, and the weights of good-performing models are increased - this is similar to a large-scale integrated learning system that is continuously trained online.

Dual scoring indicators:

• CORR (Correlation): The correlation coefficient between the prediction and the target, a direct indicator of the model's prediction ability.
• MMC (Meta Model Contribution): Measures the unique contribution of the model to the overall performance of the meta model. Even if a model has average CORR, the MMC may be high if it provides an independent signal that is different from other models. This encourages diversity in modeling strategies.

MCP and AI Agent integration: Numerai introduced the MCP protocol earlier in the field of quantitative competitions. Data scientists can use the analogy of "using AI to control the browser" to let the AI ​​Agent complete the entire process from data acquisition to model submission. The official install-mcp.sh script is provided to implement one-click mounting, allowing tools such as Claude/Cursor/Codex to directly read data, write code and submit predictions.

How to use

Numerai's usage path ranges from zero-threshold free registration to advanced MCP Agent integration, covering participants with different technical depths.

Path 1: Manual modeling (suitable for beginners)

1. Visit numer.ai/signup to register an account and obtain the API Key.
2. Download the training data set (V5.2) through the web page or API.
3. Use Python or R to train the ML model (officially provides XGBoost, LightGBM sample scripts).
4. Generate prediction files and upload them to the platform.
5. Check your CORR and MMC ratings on the leaderboard.

Path 2: Automated API calls (suitable for experienced developers)

from numerapi import NumerAPI
import pandas as pd
from xgboost import XGBRegressor

#Initialize API
napi = NumerAPI("<your_public_id>", "<your_secret_key>")

# Download data
VERSION = "v5.2"
napi.download_dataset(f"{VERSION}/train.parquet")
napi.download_dataset(f"{VERSION}/live.parquet")

#Train model
training_data = pd.read_parquet("train.parquet")
features = [f for f in training_data.columns if "feature" in f]
model = XGBRegressor(max_depth=5, learning_rate=0.01, n_estimators=2000)
model.fit(training_data[features], training_data["target"])

# Generate and upload predictions
live_data = pd.read_parquet("live.parquet")
predictions = model.predict(live_data[features])
submission = pd.Series(predictions, index=live_data.index).to_frame("prediction")
submission.to_csv("submission.csv")
napi.upload_predictions("submission.csv", model_id="your-model-id")

Path 3: MCP Agent mode (for AI native workflow)

git clone git@github.com:numerai/example-scripts.git
cd example-scripts
curl -sL https://numer.ai/install-codex-mcp.sh | bash
codex exec --yolo "find the best neural network architecture to predict target ender"

In this mode, the AI Agent automatically completes the entire process of data exploration, model architecture search, training and submission, and the data scientist only needs to describe the target in natural language.

Path 4: Numerai Signals (use own data)

If participants have their own stock market data or alternative data sources, they can submit predictive signals based on their own data through the signals.numer.ai platform, which also supports the NMR staking mechanism.

Path 5: Model Marketplace

Models developed by the community can be bought and sold through the community market. Participants cannot directly mortgage other people's models, but they can download the benchmark model as a baseline reference.

Entrance	Purpose	Free/Paid
numer.ai	Main competition platform, registration, downloading data, uploading predictions	Free
signals.numer.ai	Own data signal competition	Free
crypto.numer.ai	Cryptocurrency Competition (Entertainment/Education)	Free
docs.numer.ai	Full documentation: data, modeling, scoring, staking	Free
github.com/numerai/example-scripts	Example scripts and notebooks	Free (MIT)
forum.numer.ai	Community Forum	Free
discord.gg/numerai	Discord Community Support	Free

Product Pricing

Numerai’s pricing model is completely different from traditional SaaS – the platform itself is free for all participants, and monetization is achieved through hedge fund management fees and NMR token economics.

Data Scientist Portal: All free. There is no charge for registration, data download, model submission, ranking list viewing, and community access. NMR staking requires participants to self-purchase NMR tokens (which can be purchased on exchanges such as Coinbase), but staking is optional - you can participate in the competition and receive ratings without staking.

NMR Token Economy: NMR is the core incentive tool of the platform rather than an investment product. Participants stake NMR → NMR rewards are obtained if the model performs well → NMR is destroyed if the model performs poorly. Numerai officially states that NMR "is not a cryptocurrency", "is not an investment product", and "does not constitute ownership or control." NMR can be traded on exchanges, but there is no official guarantee of its value. As of July 2026, Numerai has conducted three strategic NMR buybacks to deepen incentive alignment with the community.

Institutional Investors: Hedge funds are open to qualified investors. The specific rates (management fees/performance compensation) are not disclosed. Please contact investing@numer.ai. The $500 million in capacity acquired by JPMorgan implies that institutional-level pricing is dominated by a "capacity allocation + performance fee" structure.

Application scenarios

• Quantitative Strategy R&D Validation: Quantitative researchers and machine learning engineers can use Numerai's high-quality financial data sets to quickly validate new predictive models and feature engineering ideas without having to solve data cleaning and label generation problems themselves. The data has been professionally cleaned and regularized and is ready for modeling. Acceptance concerns: CORR stability of the model on the validation set and uniqueness of MMC.

• Machine Learning Competition and Skills Improvement: For data scientists who want to improve their practical ML capabilities, Numerai provides a competition environment that is close to real quantitative trading scenarios, and there is no financial knowledge threshold. New data is released every two weeks, forming a continuously iterative learning cycle. Acceptance Focus: Contestants should focus on the performance stability of the model between different eras, rather than a single high score.

• AI Agent and MCP Experimental Site: Numerai’s MCP interface and AI Scientist functionality make it an ideal scenario for testing AI Agent programming and data analysis capabilities. Developers can use natural language instructions to let AI complete the entire modeling process and evaluate the automated data science capabilities of current cutting-edge models. Acceptance focus: It all depends on whether the AI ​​Agent can independently complete feature analysis, model selection, parameter adjustment and submission.

• Alternative data signal monetization (Signals): Teams with unique data sources (such as satellite images, supply chain data, consumer behavior data) can convert their alternative data into predictive signals through Numerai Signals, and obtain economic returns through the NMR mortgage mechanism without building their own transaction execution systems. Acceptance concerns: Quality score of the signal and MMC correlation with other models.

• Institutional Quantitative Allocation: For large asset managers, Numerai’s meta-model provides an alternative source of alpha with low correlation to traditional quant strategies. JPMorgan’s $500 million capacity allocation has demonstrated the value of allocation at the institutional level. Acceptance focus: The correlation between the fund’s historical risk-return characteristics and the institution’s own portfolio.

Applicable people

• Data Science & ML Practitioner: Data scientists who want to apply their machine learning skills to finance for real financial returns. Numerai provides a zero-threshold entry path and continuous data flow, which is suitable for transfer learning from Kaggle competition to quantitative direction. Not a good fit: Users with no programming background who are unfamiliar with Python/R and basic ML concepts.

• Quantitists & Hedge Funds: Quantitative teams looking for external sources of alpha or looking to develop low correlation to traditional strategies. Numerai’s metamodels can serve as a complementary signal source to existing investment research processes. Not suitable: Institutions that need full control over data sources and transaction execution processes, Numerai’s encrypted data cannot be used for their own transactions.

• AI Agent developers and researchers: developers who focus on the boundaries of AI Agent capabilities in automated data science. Numerai's MCP interface provides an end-to-end testing environment. Not suitable: A situation that has extremely high data privacy and security requirements and cannot send data to a third-party LLM.

• Cryptocurrency & Blockchain Explorer: Researchers interested in the economic design of NMR tokens and the application of blockchain in machine learning. Numerai’s Atomic Blockchain Staking and token burning mechanisms are unique case studies. Not Suitable: Traders who view NMR as a purely investment/speculative instrument - it is officially stated that NMR is not an investment.

• Academic Research and Educational Institutions: Numerai's data sets can be used for quantitative finance, machine learning teaching and academic research, and the platform has also launched a University Badges project. Not suitable: Research scenarios that require unencrypted raw financial data.

Summary and Outlook

Numerai's core competitiveness lies in the fact that it has created a three-party flywheel in which "data scientists produce models, hedge funds produce capital, and blockchain produces trust": Free encrypted data lowers the threshold for participation → A large number of model contributions improve the quality of meta-models → Fund performance attracts more institutional capital → Higher returns attract more data scientists. The 25.45% net return in 2024 and JPMorgan’s $500 million capacity endorsement prove that this flywheel has been proven feasible at the institutional level.

The current limitations are mainly as follows: (1) The fund is only open to qualified investors, and individual investors cannot directly participate; (2) The market value fluctuation of NMR tokens will amplify the actual incentive effect of participants; (3) Although encrypted data lowers the threshold of financial knowledge, it also means that the model cannot be reused in other scenarios, forming a "data lock" effect; (4) Some contents of the platform documents are still difficult for non-technical users to understand.

Procurement/Adoption Risk Assessment: For individual data scientists, Numerai is a near-zero cost skill improvement and potential source of income. The main investment is time rather than money, and the risks are controllable. For institutional investors, they need to focus on verifying the fund's historical risk and return data (performance in 2024 does not represent the future), its correlation with its own portfolio, and the specific terms of the JPMorgan capacity agreement. Before committing large amounts of capital, it is recommended to conduct smaller pilots to verify the stability of metamodel signals and complementarity with existing strategies, and to review fund documents with the assistance of legal and compliance advisors. Numerai's development direction - including atomic staking MCP Agent integration and university cooperation projects - shows that it is evolving from a single competition platform to an "AI quantification operating system" and deserves continued attention for its data set expansion, institutional cooperation progress and the maturity of the NMR token economic model.

Related tools: Hugging Face, replicate

Version Info

• Numerai Dataset V5.2 (Faith2) ：Faith2 dataset update, introducing 186 new features and two new targets (Ender and Jasper, each with 20D and 60D variants), all files are only available in int8 format. There is no official precise date yet, it is measured in months. （2025-12）
• Numerai Dataset V5.1 ：The intermediate version of the V5 series includes incremental updates and optimizations to the feature set. There is no official precise date yet. （~2025-06）
• Numerai Dataset V5.0 ：The first version of the V5 series has made major updates to the overall data architecture and introduced a new feature system. There is no official precise date yet. （~2025-01）
• Numerai Dataset V4.2 ：Final update to the V4 series, including multiple feature and target adjustments. There is no official precise date yet. （~2024-06）