Details

Autor(en) / Beteiligte

• Abrate, Carlo
• Angius, Alessio
• De Francisci Morales, Gianmarco
• Cozzini, Stefano
• Iadanza, Francesca
• Puma, Laura Li
• Pavanelli, Simone
• Perotti, Alan
• Pignataro, Stefano
• Ronchiadin, Silvia
• Kourtellis, Nicolas
• Hammer, Barbara
• Lozano, Jose A.
• Dong, Yuxiao

Titel

Continuous-Action Reinforcement Learning for Portfolio Allocation of a Life Insurance Company

Ist Teil von

• Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track, p.237-252

Ort / Verlag

Cham: Springer International Publishing

Link zum Volltext

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• The asset management of an insurance company is more complex than traditional portfolio management due to the presence of obligations that the insurance company must fulfill toward the clients. These obligations, commonly referred to as liabilities, are payments whose magnitude and occurrence are a byproduct of insurance contracts with the clients, and of portfolio performances. In particular, while clients must be refunded in case of adverse events, such as car accidents or death, they also contribute to a common financial portfolio to earn annual returns. Customer withdrawals might increase whenever these returns are too low or, in the presence of an annual minimum guaranteed, the company might have to integrate the difference. Hence, in this context, any investment strategy cannot omit the inter-dependency between financial assets and liabilities. To deal with this problem, we present a stochastic model that combines portfolio returns with the liabilities generated by the insurance products offered by the company. Furthermore, we propose a risk-adjusted optimization problem to maximize the capital of the company over a pre-determined time horizon. Since traditional financial tools are inadequate for such a setting, we develop the model as a Markov Decision Process. In this way, we can use Reinforcement Learning algorithms to solve the underlying optimization problem. Finally, we provide experiments that show how the optimal asset allocation can be found by training an agent with the algorithm Deep Deterministic Policy Gradient.