CLUSTR-1
Last updated
Last updated
CLUSTR-1 is a cutting-edge clustering algorithm developed primarily for trading applications and forms the core of our trading dapp, Hoku. Trained initially on Base and Ethereum on-chain data, CLUSTR-1 is designed to +, enabling personalized trading recommendations. Our roadmap includes expanding its capabilities to support additional networks such as Solana, BSC, and Arbitrum, opening up a broader range of use cases.
The algorithm leverages the power of unsupervised learning to identify patterns in trading behaviors without relying on predefined labels. By characterizing wallets based on risk tolerance and transaction dynamics, CLUSTR-1 delivers insights that empower traders with tailored, actionable recommendations.
CLUSTR-1 harnesses advanced clustering techniques to uncover hidden patterns in trading data, ensuring that users receive personalized and dynamic insights. One of its standout features is the development of embeddings that capture intrinsic wallet characteristics—such as age and transaction frequency—alongside comprehensive trading histories.
In addition to its robust clustering capabilities, CLUSTR-1 employs dimensionality reduction to create a three-dimensional manifold representation of the data. This 3D representation feeds directly into Hoku’s unique explorer interface, making it easy for users to visualize and interact with complex trading data in an intuitive way.
The training of CLUSTR-1 is built entirely on on-chain data, ensuring that all insights are grounded in real, immutable blockchain records. The algorithm processes detailed wallet metrics, including wallet age, transaction frequency, and complete trading histories, to derive a comprehensive view of each wallet's behavior.
By focusing on this high-integrity data, CLUSTR-1 is able to generate reliable embeddings that accurately reflect both the innate attributes of each wallet and its broader trading activity. This approach guarantees that our model remains both transparent and robust in its analysis.
While we maintain confidentiality around the specific techniques employed, the core of CLUSTR-1 revolves around generating embeddings that succinctly capture a wallet’s essential traits and trading history. These embeddings serve as the foundation for advanced clustering methods, which group wallets based on their risk tolerance and behavior patterns.
The final step in our process involves applying dimensionality reduction to transform the high-dimensional data into a three-dimensional manifold. This manifold not only simplifies the complexity of the underlying data but also powers the innovative explorer interface in Hoku, providing users with an engaging and intuitive visualization of the trading landscape.
As an unsupervised learning model, CLUSTR-1 does not conform to traditional accuracy metrics. However, we measure its effectiveness using silhouette scores, which consistently hover around 0.7, indicating strong and meaningful cluster separation.
Beyond these quantitative metrics, our backtesting has shown that many of the trading recommendations generated by CLUSTR-1 have achieved returns exceeding 100x. This impressive performance underscores the model’s practical value and its potential to revolutionize trading strategies.
The primary use case for CLUSTR-1 today is driving personalized trading recommendations within Hoku, allowing users to optimize their strategies based on nuanced insights into wallet behavior. By tailoring advice to individual risk profiles and trading histories, the model helps traders navigate the market with greater confidence and precision.
Looking ahead, the versatile nature of CLUSTR-1 positions it for expansion into other domains. Future applications could include personalized experiences in gaming, SocialFi platforms, and even the emerging market for collectible NFTs, demonstrating the wide-ranging potential of our clustering approach in the broader Web3 ecosystem.
