Learning to Manipulate a Financial Benchmark

Financial benchmarks estimate market values or reference rates used in a wide variety of contexts, but are often calculated from data generated by parties who have incentives to manipulate these benchmarks. Since the London Interbank Offered Rate (LIBOR) scandal in 2011, market participants, scholars, and regulators have scrutinized financial benchmarks and the ability of traders to manipulate them. We study the impact on market welfare of manipulating transaction-based benchmarks in a simulated market environment. Our market consists of a single benchmark manipulator with external holdings dependent on the benchmark, and numerous background traders unaffected by the benchmark. We explore two types of manipulative trading strategies: zero-intelligence strategies and strategies generated by deep reinforcement learning. Background traders use zero-intelligence trading strategies. We find that the total surplus of all market participants who are trading increases with manipulation. However, the aggregated market surplus decreases for all trading agents, and the market surplus of the manipulator decreases, so the manipulator’s surplus from the benchmark significantly increases. This entails under natural assumptions that the market and any third parties invested in the opposite side of the benchmark from the manipulator are negatively impacted by this manipulation