Forecasting and trading strategies in a market ecology

Thesis

This thesis focuses on real-world applications of a large-scale, data-driven, quarterly agent-based model of the Russell 3,000 stock index. This agent-based model, grounded in the market ecology approach, simulates 6,000 unique real mutual and exchange-traded funds. Based on holdings data from the SEC, we estimate the investment style of each simulated fund to replicate its real strategy. These funds trade in a stock market replicating the Russell 3,000 index, using historical stocks' fundamentals. After investigating the existence of a relationship between funds' trading activity and excess stock returns, we develop a novel continuous model of funds' investment styles. In our agent-based model, individual funds take deterministic trading decisions based on those investment styles estimated from holdings data. These trading decisions generate endogenous prices through market clearing. We simulate this investment universe between 2010 and 2023, initialising in each quarter the largest stocks and funds universe available from the data. We explore real-world applications of the resulting data-driven model, expanding the capabilities of current financial agent-based models. First, this agent-based model outperforms naive statistical benchmarks in forecasting quarterly excess stock returns. Our model achieves an out-of-sample accuracy score of up to 53\% in predicting excess stock returns' signs. We then backtest quarterly long-short trading strategies based on the ABM returns forecasts. ABM-based strategies display statistically significant out-of-sample Sharpe ratios and alpha when deployed in a large-cap, higher coverage ratio fraction of the Russell 3,000 universe. In this universe, ABM-based strategies display out-of-sample Sharpe ratios in a 95% confidence interval between 0.3 and 3.2, with sample estimates varying between 1.37 and 1.96, statistically significantly superior to index buy-and-hold strategies' Sharpe ratio estimates. ABM-based strategies also display a low correlation with some common existing strategies and index buy-and-hold strategies, opening the possibility of diversification and ABM-index ``smart-beta'' combined strategies. Robustness checks highlight that the source of the forecasting and trading potential of the ABM lies in our model of funds' investment styles. The relative success of those findings motivates further agent-level modelling efforts for real-world financial applications. We conclude by opening towards future research using this model for interactive backtesting of trading strategies, research on the market impact and capacity constraints of trading strategies, and a laboratory for exploring the consequences of significant changes in market ecology, such as the growth of index funds.

Authors: Vié, A

• DOI: 10.5287/ora-vjzy9k6x4
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: computational finance; agent-based modelling
• Subjects: computational finance; agent-based modelling