Pan Yiannakou and Charlie Drew, Swarm’s pioneering founders, dared to think differently and have developed a unique systemic trading program inspired by the behaviour of ants. Their Swarm XVI™ program has delivered a 60% return in its first 3 years of live trading with a Sharpe Ratio above 1 (1.35 net), won multiple industry awards and is seeing its assets under management rapidly scale.
The Swarm XVI™ program’s 2021 performance earned it the Best Performing Fund in 2021 in the AI Multi-Asset Strategy (AUM < $100m) category at The Hedge Fund Journal CTA and Discretionary Trader Awards 2022. “We expect a Sharpe of 1 is realistic over the long run, while acknowledging that recently increased market volatility could have increased the informational value of price data that generates the signals, thereby allowing us to outperform,” says Swarm CEO Pan Yiannakou.
The Swarm program is not structurally countertrend – it switches between trend, countertrend and dormant positions in individual markets – but it is clearly doing something very different.
Yiannakou previously ran up to $2.5 billion in managed futures as part of Man Group’s associated manager program starting in the late 1990s. After leaving Man Group, Yiannakou spent some years developing a technology for high-definition digital printing on leather, which was licensed to a global multi-national in 2017. Once this project was concluded, he sought a new challenge and reconnected with long-time family friend, Charlie Drew, to explore an insect-inspired trading strategy. Drew co-founded Swarm Technology with Yiannakou and is the company’s COO.
Drew has for years been researching biomimicry applications including turbine blades modelled on humpback whale fins. Drew has written on the topic, including a 2019 published paper: From Maggots to Millions: Biomimicking the Fly to Feed Humanity from its Waste in the 21st Century.
One branch of biomimicry is swarm intelligence, which has applications in the military, robotics, telecoms and communications areas. It is this natural way of thinking on which the Swarm XVI™ trading program is based. “I became fascinated with how insects share information through pheromone trails and other means of communication, swapping and sharing information and allocating tasks. Insects’ ways of organizing work avoid the behavioural greed and fear biases of mammalian society and indeed the investment world. Ants were chosen specifically as our inspiration because they are equally sized, allocate roles in colonies and share information for interchangeable functions, working towards collective welfare. Simple agents interact locally based on simple rules. This produces a remarkable intelligence amplification effect,” says Drew.
Human brains have 87 billion neurons, 3D view, effective memories, the ability to problem solve and to learn. Humans have the capability to survive and prosper as individuals. Ants only have a few hundred thousand neurons; they can only process in 2D view, have no memory, cannot learn or problem solve as individuals. They have no individual intelligence. “Ants cannot make decisions, survive or prosper alone. They require a different approach to problem solving. The Swarm XVI™ program is inspired by this mechanism for amplifying intelligence. This way of analysing information and making decisions is completely different from the traditional approach of trading rooms, which falls prey to behavioural biases,” says Yiannakou, who developed the code.
Upon returning to the investment industry, Yiannakou originally considered some form of revamped trend following, but a fascination with natural decision-making systems led him to follow this radically different framework – though a sophisticated medium term trend following indicator is the foundation, this is modulated, neutralized or reversed by signals derived from the Swarm Matrix™.
The systems might sometimes add to a winning position or might take profits on a trend following trade before any signs of the trend reversing. They can also go flat. “Some of the systems are deliberately intended to be dormant or resting, perhaps waiting for the right opportunity. In the same way, at any point in time, 20% or so of ants in a colony are resting,” says Drew.
Given that the system moves between trend and countertrend, it may make sense to use a variety of performance benchmarks in addition to traditional trend-dominated CTA indices. In common with other CTAs, the strategy is intended to provide absolute returns regardless of conventional market moves. Recently, there has been a negative equity market correlation, although this is not a design feature.
The strategy’s three-year live Sharpe Ratio, between July 2019 and June 2022, has been above its backtest, suggesting that the simulation methodology was robust with no curve fitting. Swarm takes care to avoid “optimisation”, in terms of variables such as signals or markets traded.
For example, one price datapoint per day is used. “Open, close, high and low intraday price data might add some marginal value, but it also carries the risk of curve fitting by adding parameters. Computer power makes it exceptionally easy to curve fit, especially with data that is not independent. We would rather keep parameters as fixed and/or as limited as possible,” Yiannakou explains. And the same models apply to all markets, to maintain a diversified mix of asset classes and markets rather than base allocations on performance over a certain period. Swarm currently trades 16 liquid markets, with four drawn from each of the four asset classes of equity indices, currencies, commodities and interest rates. Asset class performance attribution naturally moves around. For the first 18 months, corn was one of the worst markets but over the past year it has become one of the best. So far commodities overall have contributed most, followed by equities and currencies with interest rates detracting from returns. Some models might upsize the outperformers and downsize the underperformers, but Swarm would view this as counterproductive. “We view adding and deleting markets in response to recent performance as over-optimising, and the long run backtest proves the benefit of the four asset classes,” says Drew.
Moreover, analysing asset classes in isolation oversimplifies the interconnectedness of markets: “The interest rate markets’ conversations with the other asset classes informs other system trading parameters,” says Drew. “All markets contribute to the information, even if they do not all contribute to positive performance.”
These inter-relationships are modelled through a Swarm Matrix™ of 16 markets, which generates 65,535 connections or new data points. There are four matrices used to generate the trading signals; in effect 16 data points are converted to over 260,000 through the Swarm effect.
Two of the matrices are based on internal program data – the strategy’s position and performance – while the other two, momentum and volatility, are independent of the program. The only external data is the 16 prices sampled once a day.
The Swarm XVI™ program does not contain any correlation calculations; “Though the model is aware of historical interactions which includes correlations,” concludes Yiannakou.
There are essentially three trades: trend, countertrend or flat/zero, and the balance amongst them varies more in the short term than the long term. “Over the 15 year backtest the models were evenly split between the three positions, but there could be much more variation over 3 months,” explains Yiannakou. Incidentally, any spread trades between the markets are accidental: “The combination of trades might sometimes look like an arbitrage or relative value trade, but this is not intentional,” he adds.
The number of markets traded is relatively small, partly because the nature of the models means that a larger matrix would require gargantuan computer power. “We run seven million calculations a day to generate a set of trades across the 16 markets we trade,” says Yiannakou. “Increasing the number of markets increases the number of connections exponentially; at 32 markets it would run into hundreds of millions and at 64 it would be trillions of calculations for which we would need NASA level computer power,” explains Drew.
Therefore, it is most likely that Swarm will expand its investment universe through parallel programs with partial overlaps as markets are added. “In an ant colony one million ants do not all talk to each other but connect directly with their neighbours. The information spreads via small groups like a Mexican wave in a sports stadium. Our current research is based on parallel expansion along with improved risk management and trading efficiency,” explains Yiannakou.
Though the system carries out a great many calculations daily, these however feed into a small number of trades. “The Swarm XVI™ program runs calculations once per day, but using a time sequence series, so that the longest lookback is one year and weightings within the matrix change quite slowly based on rolling one year data. Risk management uses a 34-day volatility lookback to try and equalize the dollar risk of the 16 markets traded,” says Yiannakou. Daily trades are more often position size tweaks than binary shifts in direction: “The average holding period is 17 or 18 days defined by the direction of a trade, but the position size can change more often with volatility and multiplier data,” he explains.
With 3,000 round turn trades per million dollars a year, trading costs are kept low. Slippage is controlled by seeking high volume trade times. “Algorithmic execution, and staggered trading multiple times per day, could be added as assets grow and when the need arises,” says Drew.
The Swarm XVI™ program has elements of artificial intelligence; a ‘network’, weightings, parameters that change based on past results. But it is not a fast mechanism with a feedback loop. There is non-specific pattern recognition.
“There is machine learning… but it is very, very slow in comparison to what is normally understood to be ML. It might be more accurate to say the program ‘evolves’ rather than learns. You could say the Swam XVI™ program is an example of Machine Evolution (ME) rather than ML,” says Yiannakou. “We are not optimizing based on previous patterns. Our weightings change much more slowly than traditional AI. We are not trying to mimic human intelligence. It is rather a completely different way of thinking and decision making,” underscores Drew. “This is how an ant colony survives. Through very gradual adaptation. Human intelligence is orders of magnitude faster, which may in the end be as much of a disadvantage than an advantage. After all, insects have been around for many millions of years longer than humans and may well survive long after we are gone,” adds Drew.
There are four layers of risk management. Level one is the internal adjustments of positions on a weekly basis. The additional levels are designed to protect assets against market dislocations and extreme events.
The second layer sees the model mechanically de-gear down to 75%/50%/25% exposure about 6 times a year, for up to a week, when various indicators are outside normal historical ranges. “The dashboard of indicators, which could include absolute and relative levels of price action and volatility, draws some inspiration from Drew’s pilot’s license, which he obtained before his driver’s license,” points out Yiannakou.
The third layer of risk management might reduce risk on an intraday basis. “Discretionary risk management simply pre-empts stop losses and other risk management triggers by exiting before stops are reached. This pre-empts other measures for a few hours, and is not designed to add performance,” says Yiannakou.
The fourth level of risk management is for individual users to determine. Hard stops are agreed with clients, who also determine their preferred level of leverage and trading account sizes. Margin stress tests assume peak margin to equity at 35% of notional.
Assets under management are currently held in managed accounts and a Cayman domiciled fund is planned for launch in 2023.
The strategy went live with an initial family office allocation in July 2019, before being opened to external investors. Now entering its fourth year of live trading it has rapidly attracted assets under management in 2022 and now has assets of over $64 million, having garnered mandates from US, European and UK institutions. Capacity is estimated at $1 billion plus.
Most operational functions are outsourced to Privium’s platform, which has over 30 clients and several billion of assets (including some managers who have featured in previous editions of Tomorrow’s Titans). There are currently three principal brokers, in Europe and the US.
Swarm Technology uses a fundamentally different way of thinking. Whereas AI tries to mimic mammalian intelligence and problem solving, Swarm takes a very different insect-based approach. The financial markets in 2022 have utterly upended some strategies that flourished for decades before. Perhaps the time is right for a new way of thinking in these turbulent times.
Past performance does not predict future results and the capital value of investments and the income generated can fluctuate.