Algorithmic trading, financial stability and market competitiveness

The thesis is composed of two chapters and investigates the impact of artificial intelligence (AI) traders in financial markets. In the first paper, coauthored with Francesco Sangiorgi and Emanuele Tarantino we whether AI traders, governed by a deep deterministic policy gradient (DDPG) algorithm learn to trade. We design a demand-based asset pricing model populated by one or more AI traders and a representative investor, whose demand is calibrated exploiting investors-holding data from 13F SEC filings and following Koijen and Yogo (2019). Equilibrium prices are determined as function of public signals (i.e. stock characteristics), changes in the unobserved latent demand. Further, AI traders contribute to price formation through their price impact. We compare the the portfolio choices of AI traders with the optimal policy of a rational expectation benchmark where investors know the price formation process and internalize not only their own price impact, but also the price impact of their competitors. Last, we study how the investment behaviour of AI traders shape market efficiency and liquidity. In our simulations we show that when the number of AI trader is limited to one it learns to exploit return predictability, decode latent demand from public information and adjust its strategy to account for its price impact. Quantitative differences emerge when multiple agents populate the market, with discrepancies between AI traders’ policy and the portfolio choice of rational expectation investors rising as competition increases. This negative learning externality deteriorates trading profits and market efficiency compared to the rational benchmark. In the second paper, I contribute to the recent literature studying market elasticity and performance by investigating the impact of artificial intelligence (AI) traders on aggregate stock elasticity. Building on Haddad et al. (2021), I simulate a financial market exhibiting realistic returns populated by one representative investor, whose demand for assets is microstructured and calibrated on investors-holding data, and one AI trader. In this framework, both asset prices and aggregate elasticity are endogenously determined in equilibrium by the portfolio choices of the representative investor and AI trader. First, I report that the AI trader has a demand elasticity four times larger than the average US institutional investor. However, the presence of more active investors is only a necessary but not sufficient condition to increase market elasticity. On the one hand, this agent trades actively and thus is a candidate for increasing elasticity. On the other hand, whether market elasticity increases depends on the allocation chosen by the AI trader. I show that when the agent actively participates in the market, aggregate elasticity increases compared to a benchmark case where the AI trader is absent. However, exactly because of its trading aggressiveness, it withdraws liquidity from the market, further enhancing the negative trend in aggregate elasticity. viii