To model the real world, quants turn to synthetic data
27 April 2020
Black swans are like kryptonite to quants. When faced with extreme and unexpected events, their carefully constructed models seem to quickly lose their predictive power. The coronavirus pandemic is another example of this. With consumers locked down and businesses in limbo, markets saw some of the largest daily swings in nearly a century. Quant models struggled to cope…
Simudyne has modelled price formation processes in equity markets, including replicating different exchange protocols, such as how limit and market orders are handled. According to the firm, its simulated trading environment – built on nanosecond-level tick-by-tick data from Nasdaq and the London Stock Exchange’s Turquoise platform – produces synthetic data that is statistically indistinguishable from real market data.
Data vendor Refinitiv is experimenting with Simudyne’s software to create packages of synthetic data for sale. Geoff Horrell, director of applied innovation at Refinitiv’s London Lab, says it chose ABM partly because, “intuitively, people understand the concept of an agent”.
The datasets are being marketed to banks for use in testing execution algorithms. “Banks need their algos to operate effectively under different market conditions,” says Justin Lyon, chief executive of Simudyne.
“Being able to dial up extreme volatility or very thin liquidity is something they’re interested in. Let’s say you have an event where liquidity is drying up. You might want to know the correlated behaviours that led to a market crash, and which types of traders were trading more with other types of traders…”
Simudyne’s current model is largely parametric, but also mixes in some non-parametric elements, according to Lyon: “We will always have this element where we build the behavioural logic of agents, but we also use an empirical distribution to drive our simulation, so that we can catch things like heavy tail, which would not be traditionally possible if we had some purely parametric distribution of active volume.”
In its research, the company has begun exploring the creation of a new class of agent that, unlike traditional agents who make trading decisions based on rules, utilises recent advances in deep learning and employs different trading objectives. Horrell says initial results confirm the proposed method is superior to classical techniques and “absolutely is the next level of future work in this area”.
To continue reading please visit Risk.net.