Technology and trading

Chowdhury studied computer science including a Masters at UCL and the other team members also have Masters or Doctorates in IT-oriented subjects. They previously worked for Goldman Sachs, Credit Suisse, Barclays, UBS and BNP Paribas, in London, Luxembourg and New York. Chowdhury started as a front office software developer in quantitative finance commodities, pivoted to trading and became head of Deutsche Bank’s commodity derivatives desk in London, running just under USD 1 billion in metals and energy in the derivatives space. “This was a mix of flow-driven market making and some proprietary trading of about USD 100 million subject to rules and limited risk regulations with a strict annual profit target,” says Chowdhury.

A home-grown garage hedge fund

In 2017 Chowdhury and Anthony Silver started researching machine learning and AI in more detail and Chowdhury persuaded his wife to let him start trading a pool of capital with Silver, whose partner also agreed to the enterprise. “We were confident that the strategy would work across asset classes and deliberately designed algorithms that were not specific to any asset class. Though my background had been in commodities, it was simpler to apply the ideas to equities because they are easier to access without an ISDA license,” says Chowdhury.

Whereas many start-up managers join a giant multi-strategy prop shop and trade at least USD 10-50 million, AVO was started with personal capital. “We took some pride in starting this fund in the garage and basement with a small amount of capital. We were trading USD 1-2 million from personal pensions and other liquid assets; we owned 100% and had total control over research. We were not under any pressure to rush the research or subject to any deadlines and found that more slow and steady development was more robust,” recalls Chowdhury. The 7-year track record since 2017 contains a few gaps since there was no need to be fully invested and they sometimes spent a few weeks or months figuring out new strategies; the models have been regularly evolving by adding new data weekly or monthly. Trading has been continuous since Ancova launched the flagship strategy in 2021. Chowdhury and Silver remain the key quants driving algorithm development and are also authorised for trading and execution. Andrew Bradley and Simon Kaufman mainly handle IT, back office and infrastructure.

“This is very much a team effort, and no one is indispensable,” stresses Chowdhury. “A key difference between us and other hedge funds is cross-pollination of ideas into a unified strategy. Other funds have multiple desks, each dealing with their own books and strategies, and the purpose is to ensure risk reduction by having successful desks outperform unsuccessful ones. That also reduces sharing of information. At Ancova if a technique is shown statistically to increase alpha it is incorporated into the main strategy so all teams benefit,” explains Chowdhury.

Proprietary option pricing techniques

Chowdhury spent years honing and refining improvements to traditional option pricing algorithms. “We knew that the bank algorithms based on models only worked in idealised circumstances if certain factors stay constant. We started by looking at the Heston model, which was viewed as being state of the art. Heston is a beautiful mathematical theory, but it assumes volatility and returns are fixed, and does not include correlation, beta or theta. We identified its faults and wrote our own volatility and option pricing models improving Heston with more factors. We also allow for non-normal distributions unlike Black Scholes. Whereas quantitative mathematicians must prove theories with a closed form solution, as engineers we do not have to prove them in that way. We just expect to prove an edge and in 2017 we ran our models and got a much better level of risk adjusted returns,” says Chowdhury.

From hedge to profit centre

In 2019 the managers started using options only to hedge downside risk but later realized that their insights into option mispricing could become a profit centre. “We developed our own machine learning algorithm, customised to our needs, to figure out the volatility surface,” recalls Chowdhury. Options therefore play a dual role: “This is not exactly volatility arbitrage, but we are seeing mispriced options. The options are both part of the investment strategy and a risk management overlay. Some of our options are expected to lose money. Very occasionally we may hold a single option without the stock,” says Chowdhury.

Proprietary AI optimises mixes of equities and options

High performance computing generates a machine learning model to construct an optimised portfolio. “We are not predicting the direction of assets but rather optimising for combinations of equities and options that produce the right profit and loss profile,” explains Chowdhury.

The AI-powered portfolio optimiser improves on Markowitz, which assumes infinite amounts of capital and linear options. “Our optimiser works for finite capital and nonlinear options,” says Chowdhury. The optimiser can select stocks from the S&P 500 index but only chooses its favourite 150 before further diversification filters arrive at 30-40 stocks. “We can capture 80% of upside and downside without all 500 stocks,” points out Chowdhury.

“The portfolio has low downside beta but not market neutral – the options provide the risk hedging while preserving returns,” he continues. The beta of stocks held is between 0.6 and 1.2 but options are also key to reducing portfolio volatility. Of 35 or 40 stocks about 65% or 70% are long positions and 30% or 35% are short positions.

Some stocks have an option hedge in the same stock, others may have option cross hedges on different stocks, and others have no option hedge at all. The optimiser could simply sell calls and buy puts on single stocks, or it might trade a wide variety of option spreads using 1×1 ratios or other ratios. It can also put on more complicated calendar spread structures.

The strategy trades correlations amongst both stocks and options. “Correlations between options are different from those between stocks,” says Chowdhury.

In addition, the model sells overpriced and buys underpriced options. This is dubbed “statistical arbitrage”; this is clearly very different from the traditional description of “statistical arbitrage” and would be perceived as “volatility arbitrage” by most allocators.

“Rebalancing no more often than monthly reduces transaction costs and results in very stable portfolios,” says Chowdhury. Options must be rebalanced monthly, and stocks can also be reshuffled at the same time but may be rebalanced somewhat less often. Al execution is electronic and algorithmic, though there is no direct market access.

“We could not do all of this without our own AI. We use supervised deep learning neural networks and matrix algebra where appropriate. I still use 80% of the maths and matrix algebra from my Masters in AI but now we also have high performance engineering (HPE) expertise providing an edge. HPE is more widely used in games, but I wrote the world’s fastest Black Scholes pricer using HPE techniques to exploit the benefits of 2020s computers,” reveals Chowdhury.

Performance drivers

The portfolio lagged the S&P 500 in 2023 and 2024 when the MAG 7 were driving returns but has also shown better downside protection in down periods. The aim is to make a steadier 12-20% return, which was attained in both of the past two years. “The options are expected to make an average of 1% a month, though they will not all profit. There might be 15 positive and 15 negative option positions. If all our options were positive that would imply that we had no protection. In a small up market, option costs can outweigh stock gains. It is in a sharp down market when options really come into play,” explains Chowdhury.