A pair of researchers at the University of Tsukuba in Japan recently built an artificial intelligence-powered cryptocurrency portfolio management system trained using on-chain data, the first of its kind, according to scientists.

The artificial intelligence (AI) system, called CryptoRLPM, short for “Cryptocurrency Reinforcement Learning Portfolio Manager,” utilizes a training technique called “reinforcement learning” to implement on-chain data into its models.

Reinforcement learning (RL) is an optimization paradigm in which an AI system interacts with its environment (in this case, a cryptocurrency portfolio) and updates its training based on reward signals.

CryptoRLPM applies feedback from RL throughout the architecture. The system consists of five main units that work together to process information and manage structured portfolios.

These modules include data feed unit, data refinement unit, portfolio proxy unit, real-time transaction unit and proxy update unit.

Screenshot of a preprint study. Source: Huang, Tanaka, “A Scalable Reinforcement Learning System for Cryptocurrency Portfolio Management Using On-Chain Data”

Once developed, the scientists tested the CryptoRLPM by allocating three portfolios. The first contains only Bitcoin (BTC) and Storj (STORJ), the second keeps BTC and STORJ while adding Bluzelle (BLZ), and the third keeps all three plus Chainlink (LINK).

These experiments were conducted between October 2020 and September 2022 and were divided into three distinct phases (training, validation, and backtesting).

The researchers measured CryptoRLPM’s success against a baseline assessment of standard market performance through three metrics: Accumulated Return Rate (AAR), Daily Return Rate (DRR), and Sortino Ratio (SR).

AAR and DRR are intuitive measures of an asset’s loss or gain over a given time period, and SR measures an asset’s risk-adjusted return.

Screenshot of a preprint study. Source: Huang, Tanaka, “A Scalable Reinforcement Learning System for Cryptocurrency Portfolio Management Using On-Chain Data”

According to the scientists’ preprint research paper CryptoRLPM exhibit Significant improvement over baseline performance:

“Specifically, CryptoRLPM improves ARR by at least 83.14%, DRR by at least 0.5603%, and SR by at least 2.1767 compared to the benchmark Bitcoin.”

related: DeFi meets AI: Could this synergy be the new focus for tech acquisitions?