Risk averse asset allocation in a context of climate change with reinforcement learning and hidden Markov models

We aim to address the challenge of identifying an ideal long-term arbitrage strategy that can adapt to an individual's market perspective. In this research, we expand upon the existing body of knowledge regarding optimal asset allocation by employing a reinforcement learning algorithm rooted in a Markov Decision Process. The state space of this process is defined by estimating a Hidden Markov Chain (HMC), which serves to characterise the market dynamics. Our agent acquires knowledge about the market at each time step through this characterisation, employing the MAP algorithm to determine an optimal strategy. We then extend the agent's state space to incorporate a physical risk index and a climate transition risk index. After showing the current limits to the integration of such indices, we explore three possible scenarios for the materialisation of climate risks on market regimes, which could be integrated into climate stress tests, and analyse the behaviour of our agent in each of these scenarios. This article therefore illustrates the relevance of estimating an HMC to construct resilient allocation strategies that could be used in climate stress tests that make assumptions about the impact of climate change on volatility.