Syvävahvistusoppiminen keskitettyjen likviditeettipositioiden suojaamisessa

Abstract

Decentralized finance has introduced new forms of market making through automated market makers, where users provide liquidity to decentralized exchanges such as Uniswap. In Uniswap v3, liquidity providers (LPs) can concentrate their liquidity within custom price ranges, improving capital efficiency but introducing exposure to impermanent loss and nonlinear portfolio risk. Managing this exposure dynamically poses a challenge as the LP position’s token composition changes continuously with market prices. This thesis investigates whether a Deep Reinforcement Learning (DRL) agent can effectively hedge a Uniswap v3 LP position using cryptocurrency futures. A simulated Uniswap environment is developed to model concentrated liquidity providing positions and the hedge positions tied to it. The hedging problem is framed as a sequential decision process, where the agent seeks to minimize downside portfolio variance while maintaining upside exposure. The Proximal Policy Optimization algorithm is applied to learn an adaptive hedging policy, which is evaluated against baseline strategies such as fixed-frequency rebalancing. The results show that the DRL-based hedges can outperform the baseline strategies, achieving higher average portfolio returns, with similar average drawdowns. However, the learned policies varied between agents, and although they reduced downside variance in many cases, they were also more likely to experience larger maximum drawdowns. These findings indicate that reinforcement learning offers a promising but complex approach for managing the nonlinear risks of concentrated liquidity provision in decentralized exchanges.