Gaining insight on future results of trading strategies is not just about back-testing algorithms against historical market data. Predicting future price movements requires modeling a wide range of market conditions to explore what might happen.
- A Refinitiv white paper examines a model for simulating financial markets through a combination of agent-based modeling, synthetic data and machine learning.
- Agent-based modeling observes the collective behavior between large numbers of autonomous agents — high frequency, fundamentalist and chartist traders.
- The research presents agent-based simulation models built on Simudyne software, highlighting the ability to generate price paths for various ‘what-if’ scenarios.
Financial professionals extensively use historical market data in order to gain insight into the effectiveness of their trading strategies.
The presumption is that the market comprises recurring patterns and by studying these patterns in the past, one can predict future price movements.
There are several limitations with this approach. Relying on the belief that future events can be calculated with actuarial certainty from past data is deeply flawed.
It is important to consider all possible outcomes, including those that are outside of historical bounds, for efficient modeling of uncertainty and to avoid the danger of overfitting.
Simulating financial markets
Computational simulations are an effective mechanism to augment historical data. They can model a wide variety of market conditions and explore what might happen under extreme situations.
Classic simulation techniques that take a top-down modeling approach are not suitable because the dynamics of financial markets are just too complicated to be represented by structural models.
A recently-released white paper by Refinitiv — Synthetic Reality: Synthetic market data generation at scale using agent based modeling — explores new ways of simulating financial markets by combining three technologies:
- Agent-based modeling (ABM)
This has been developed as a tool of last resort, to obtain results when a phenomenon that is to be modeled is too complex for traditional approaches.
An ABM takes a bottom-up approach and may more realistically capture the complex dynamics of financial markets. It studies the interactions between large numbers of individuals termed agents, which possess independent decision-making capabilities.
It has been used in the past for simulating the interaction between military powers in the Cold War, or among societies and biological ecosystems, as well as for simulating financial markets.
Our approach relies on the Simudyne platform, which ensures simulation can scale beyond tens of thousands of parallel agents. Our experiments combine Refinitiv’s mathematical models implemented in Python with Simudyne’s Java-based simulation software.
- Synthetic data
Historical data as the sole source of modeling focuses too narrowly on what has happened, and cannot provide support for answering the question of what could happen.
Random data that has been generated to exhibit certain properties of financial markets may be random but it is also plausible, and can therefore serve to supplement historical data for the analysis of scenarios that go beyond the historical past.
- Deep learning
Traditionally each agent in ABM has a way of acting that is hardwired in the form of rules. More recently, researchers have used machine learning inside individual ABM agents to make their behavior more adjustable by the state of the simulation (i.e., the environment).
The approach in this white paper shows for the first time that deep learning can be used successfully as an ABM agent’s action strategy.
Across multiple asset classes
The white paper shows how thousands of traders with different strategies — high frequency, fundamentalist, chartist — interact, leading to complex overall system behavior, and how synthetic data can be used to explore bullish, bearish and flash crash scenarios.
It also shows how synthetic and historical data can be used together with agents equipped to adapt to their environment using deep learning.
We conducted evaluations on multiple asset classes across a portfolio of assets and found that the proposed agent decision mechanism outperforms other techniques. Our simulation model also successfully replicates the empirical stylized facts of financial markets.