In this paper, we introduce DeepTraderX (DTX), a simple Deep Learning-based trader, and present results that demonstrate its performance in a multi-threaded market simulation. In a total of about 500 simulated market days, DTX has learned solely by watching the prices that other strategies produce. By doing this, it has successfully created a mapping from market data to quotes, either bid or ask orders, to place for an asset. Trained on historical Level-2 market data, i.e., the Limit Order Book (LOB) for specific tradable assets, DTX processes the market state $S$ at each timestep $T$ to determine a price $P$ for market orders. The market data used in both training and testing was generated from unique market schedules based on real historic stock market data. DTX was tested extensively against the best strategies in the literature, with its results validated by statistical analysis. Our findings underscore DTX's capability to rival, and in many instances, surpass, the performance of public-domain traders, including those that outclass human traders, emphasising the efficiency of simple models, as this is required to succeed in intricate multi-threaded simulations. This highlights the potential of leveraging "black-box" Deep Learning systems to create more efficient financial markets.

I introduce PRZI (Parameterised-Response Zero Intelligence), a new form of zero-intelligence trader intended for use in simulation studies of the dynamics of continuous double auction markets. Like Gode & Sunder's classic ZIC trader, PRZI generates quote-prices from a random distribution over some specified domain of allowable quote-prices. Unlike ZIC, which uses a uniform distribution to generate prices, the probability distribution in a PRZI trader is parameterised in such a way that its probability mass function (PMF) is determined by a real-valued control variable s in the range [-1.0, +1.0] that determines the _strategy_ for that trader. When s=0, a PRZI trader is identical to ZIC, with a uniform PMF; but when |s|=~1 the PRZI trader's PMF becomes maximally skewed to one extreme or the other of the price-range, thereby making its quote-prices more or less urgent, biasing the quote-price distribution toward or away from the trader's limit-price. To explore the co-evolutionary dynamics of populations of PRZI traders that dynamically adapt their strategies, I show results from long-term market experiments in which each trader uses a simple stochastic hill-climber algorithm to repeatedly evaluate alternative s-values and choose the most profitable at any given time. In these experiments the profitability of any particular s-value may be non-stationary because the profitability of one trader's strategy at any one time can depend on the mix of strategies being played by the other traders at that time, which are each themselves continuously adapting. Results from these market experiments demonstrate that the population of traders' strategies can exhibit rich dynamics, with periods of stability lasting over hundreds of thousands of trader interactions interspersed by occasional periods of change. Python source-code for the work reported here has been made publicly available on GitHub.

Optimal execution is a sequential decision-making problem for cost-saving in algorithmic trading. Studies have found that reinforcement learning (RL) can help decide the order-splitting sizes. However, a problem remains unsolved: how to place limit orders at appropriate limit prices? The key challenge lies in the "continuous-discrete duality" of the action space. On the one hand, the continuous action space using percentage changes in prices is preferred for generalization. On the other hand, the trader eventually needs to choose limit prices discretely due to the existence of the tick size, which requires specialization for every single stock with different characteristics (e.g., the liquidity and the price range). So we need continuous control for generalization and discrete control for specialization. To this end, we propose a hybrid RL method to combine the advantages of both of them. We first use a continuous control agent to scope an action subset, then deploy a fine-grained agent to choose a specific limit price. Extensive experiments show that our method has higher sample efficiency and better training stability than existing RL algorithms and significantly outperforms previous learning-based methods for order execution.

The logarithmic market scoring rule (LMSR), the most common automated market making rule for prediction markets, is typically studied in the framework of dealer markets, where the market maker takes one side of every transaction. The continuous double auction (CDA) is a much more widely used microstructure for general financial markets in practice. In this paper, we study the properties of CDA prediction markets with zero-intelligence traders in which an LMSR-style market maker participates actively. We extend an existing idea of Robin Hanson for integrating LMSR with limit order books in order to provide a new, self-contained market making algorithm that does not need “special” access to the order book and can participate as another trader. We find that, as expected, the presence of the market maker leads to generally lower bid-ask spreads and higher trader surplus (or price improvement), but, surprisingly, does not necessarily improve price discovery and market efficiency; this latter effect is more pronounced when there is higher variability in trader beliefs.

We report on results from experiments where human traders interact with software-agent traders in a real-time asynchronous continuous double auction (CDA) experimental economics system. Our experiments are inspired by the seminal work reported by IBM at IJCAI 2001, where it was demonstrated that software-agent traders could consistently outperform human traders in real-time CDA markets. IBM tested two trading-agent strategies, ZIP and a modified version of GD, and in a subsequent paper they reported on a new strategy called GDX that was demonstrated to outperform GD and ZIP in agent vs. agent CDA competitions, on which basis it was claimed that GDX "...may offer the best performance of any published CDA bidding strategy.". In this paper, we employ experiment methods similar to those pioneered by IBM to test the performance of "Adaptive Aggressive" (AA) algorithmic traders. The results presented here confirm Vytelingum's claim that AA outperforms ZIP, GD, and GDX in agent vs. agent experiments. We then present the first results from testing AA against human traders in human vs. agent CDA experiments, and demonstrate that AA's performance against human traders is superior to that of ZIP, GD, and GDX. We therefore claim that, on the basis of the available evidence, AA may offer the best performance of any published bidding strategy.