News & Events

Artificial Intelligence: The New Frontier for Hedge Funds (2/2)

This piece revisits the performance of hedge funds that utilise artificial intelligence and machine learning theory in their trading process, focusing on the overall risk-return profile of artificial intelligence (AI) hedge funds as captured by the Eurekahedge AI Hedge Fund Index in comparison to traditional quants, equity-hedge strategies and the average global hedge fund. We conclude with an interview with Mr. Yoshinori Nomura of Simplex Asset Management who shares the nuts and bolts about his pure AI strategy and Mr. Atsuo Ogaki and Mr. Damien Loh of Ensemble Capital who discuss their AI strategy which follows the ‘Man and Machine’ approach. Both funds conclude with their respective views on the outlook for AI application in the hedge fund space and how it could provide that edge over traditional quant strategies.

Quantitative hedge fund strategies have received considerable interest from investors over the last decade. The application of growing computing power and the availability of big data has enabled these systematic trading models to capitalise on market inefficiencies that were otherwise difficult to identify or harvest given the implied trading costs. However, this growth has met with some headwinds on two key accounts; firstly, trading models built using back-tests on historical data have often failed to deliver good returns in real time (as previously identified trends have broken down), and secondly, the diffusion of similar quant models which has led to crowding in the space and consequently depressed the returns from such strategies. The latter of the two problems is being addressed to some extent through the use of even faster processing machines (competition in the high frequency trading space) as funds race to place their trades ahead of the crowd. However, it is the developments on the first front that could transform the quant landscape as trading models upgrade from identifying trends (through back tests) to ‘predicting and adapting’ to the trends in real time (such as using walk forward tests) – i.e. the application of artificial intelligence and machine learning theory to hedge fund quants.

The application of AI in the hedge fund industry is still at an early stage albeit growing by the day – with some hedge fund managers utilising AI as a partial input into their trading process (retaining their discretionary control over investing and risk management) whilst others, what we call ‘pure AI hedge funds’, who have outsourced both the trading and risk management aspect to the machine with minimal input from the fund manager. This variability in the application of AI in the hedge fund space mirrors its development in other industries as well; such as transport sector where a fully reliable self-driven car is yet to hit the roads – and the challenge will lie in successfully showcasing the ability of the machine in learning meaningfully, and in the context of hedge funds, consistently and profitably for their investors.

Figure 1: AI Hedge Fund Index vs. quants and traditional hedge funds
AI Hedge Fund Index vs. quants and traditional hedge funds

Figure 1 compares the Eurekahedge AI Hedge Fund Index with traditional quants (denoted by the Eurekahedge CTA/Managed Futures Hedge Fund Index and it’s sub-group of systematic trend followers), traditional hedge funds as captured by the Eurekahedge Hedge Fund Index and the Eurekahedge Long/Short Equity Hedge Fund Index. The AI hedge fund index tracks 28 funds historically (including dead ones to offset any survivorship bias in index values) and as of present tracks the performance of 17 live actively trading funds. As can be seen in Figure 1, AI hedge funds have outperformed both traditional quants and the average hedge fund since 2010, delivering annualised returns of 10.16% over this period compared with 3.31%, 2.17%, 5.25% and 6.11% for CTA’s, trend-followers, the average global hedge fund and long/short equity hedge funds respectively. In the difficult markets of 2016, AI hedge funds were up 11.02%, ahead of the average global hedge fund which gained 4.61% and their peers as displayed in Table 1. In 2017, AI hedge funds outperformed the average global hedge fund and their traditional quant peers (both CTAs and sub-group systematic trend following hedge funds), delivering 10.16% whilst underperforming equity hedged strategies which returned 12.63% following a strong rally in underlying equity markets. However, as the Trump rally finally came to a close earlier this month, it would be interesting to see how AI hedge funds perform going forward as some semblance of volatility returns to the markets.

Table 1 summarises the key performance statistics for AI hedge funds relative to their peers. Key takeaways include:

  1. AI hedge funds have outperformed the average global hedge fund for all years excluding 2012 when they lost 1.62% compared with a 7.36% gain for the latter.
  2. Barring 2011 and January 2018, returns for AI hedge funds have outpaced those for traditional CTA/managed futures strategies while underperforming systematic trend following strategies for the year 2014 (and January 2018 – strong directional moves in underlying markets until the correction seen in early February) when the latter realised strong gains from short energy futures.
  3. Over both the five, three and two year annualised period, AI hedge funds have outperformed both traditional quants and the average global hedge fund delivering annualised gains of 11.83%, 12.80% and 9.97% respectively over these periods.
  4. AI hedge funds have also posted better risk-adjusted returns over the last five and two year annualised periods compared to all peers depicted in the table below, with Sharpe ratios of 2.76 and 2.57 over both periods respectively.
  5. While returns have been more volatile compared to the average hedge fund (compared with the Eurekahedge Hedge Fund Index), AI funds have posted considerably lower annualised volatilities compared with systematic trend following strategies.
  6. On a two year annualised basis, AI hedge funds have delivered higher annualised returns and better risk-adjusted returns compared to traditional quants, but have lagged behind long/short equity hedge funds. The Eurekahedge AI Hedge Fund Index is up 9.97% on a two year annualised basis, versus 11.55% for long/short equity hedge funds, with the latter benefitting strongly from the one way Trump rally in 2017 which saw equity markets climb to all-time highs.

Table 1: Performance in numbers – AI Hedge Fund Index vs. Quants and Traditional Hedge Funds

Eurekahedge AI Hedge Fund Index
Eurekahedge CTA/Managed Futures Hedge Fund Index
Eurekahedge Trend Following Index
Eurekahedge Hedge Fund Index
Eurekahedge Long Short Equities Hedge Fund Index
January 2018 year-to-date
5 year annualised returns
5 year annualised volatility
5 year Sharpe Ratio (RFR=1%)
3 year annualised returns
3 year annualised volatility
3 year Sharpe Ratio (RFR=1%)
2 year annualised returns
2 year annualised volatility
2 year Sharpe Ratio (RFR=1%)

Source: Eurekahedge

Table 2 shows the correlation matrix of AI hedge funds with respect to their peers identified. It is interesting to note the strategy is negatively correlated to the average hedge fund (-0.129), traditional equity hedged strategies (-0.169) and has a small positive correlation to CTA/managed futures and trend following strategies, highlighting the positive gains that can be accrued to the overall portfolio from a potential diversification across this strategy. However, it is pertinent to note that only the correlation value versus the systematic trend following strategies is statistically significant at 90% confidence level.

Table 2: Correlation Matrix
AI and other hedge funds correlation matrix

Figure 2 depicts the 12-month rolling alpha (excess returns assuming zero risk-free rate) of AI hedge funds relative to CTA/managed futures strategy (approximating quants) and the average global hedge fund. Alpha has generally been positive with 12-month alpha as of January 2018 coming in at 0.64% and 0.14% relative to CTA/managed futures strategies and the average hedge fund respectively. As of end-December 2016, the 12-month rolling alpha value came in at 0.81% versus CTA/managed futures hedge funds and 1.10% versus the average global hedge fund, pointing perhaps towards a combination of the improved capability of AI models and the over-crowding (shrinking alpha) in comparable strategies.

Over the full period under study starting from 2011, AI hedge fund managers have delivered alpha values of 0.78%, 0.80%, 0.89% and 0.88% versus CTA/managed futures, systematic trend-followers, the average global hedge fund and traditional long/short equity hedge fund strategies.

Figure 2: 12-month rolling alpha relative to quants and traditional hedge funds
12-month rolling alpha relative to quants and traditional hedge funds

To conclude, Table 3 below looks at the performance of AI hedge funds and their peers during periods of market stress in the recent past and gauges the ability of different strategies to deliver positive returns in volatile times. Key takeaways include:

  1. AI hedge funds have been in the red in three out of the 12 events identified in the table below, the least among the group. In comparison, the average global hedge fund posted losses in six out of these 12 periods.

  2. With regards to some of the idiosyncratic risk events identified in the table below, such as the Trump win and Brexit, AI funds have posted a mixed record but have done relatively better compared to their peer group. An improvement in AI application and its learning capabilities could perhaps yield even better results in the future, but as their recent losses in November 2016 (Trump win) show, there are clear limits to predicting the future for the machines.

  3. Interestingly however, AI hedge funds have posted compound returns of 13.73% in the Trump rally which ended in February this year, second only to the returns for equity long/short hedge funds and almost twice that when compared to traditional quants and the average global hedge fund. This seems to suggest that AI hedge funds have been able to adjust to the lower market volatility environment quite successfully which generally mutes returns for traditional quant hedge funds.

Table 3: AI hedge fund returns during key market risk events

Eurekahedge AI Hedge Fund Index
Eurekahedge CTA/Managed Futures Hedge Fund Index
Eurekahedge Trend Following Index Eurekahedge Hedge Fund Index Eurekahedge Long Short Equities Hedge Fund Index
Dec 2016 To Jan 2018
(Cumulative Return)

Trump Rally Post Nov-2016

Nov 16
Trump Win
Jun 16
Feb 16
Oil Price Dip/China growth concerns
Jan 16
Oil Price Dip/China growth concerns
Aug 15
China Equity Crash
Jul 15
China Equity Crash
Jun 15
Greek referendum
Jan 15
Swiss Franc De-pegging
Sep 14
Oil Price Dip
Jun 13
Taper Tantrum
May 13
Taper Tantrum

Source: Eurekahedge

Interview with Yoshinori Nomura, Director at Simplex Asset Management Co., Ltd.

Simplex Asset Management Co., Ltd. is an independent hedge fund and investment management firm in Japan founded in 1999. The AUM is US$430 billion as of December 2017.

  1. Simplex Equity Futures was first featured in the Eurekahedge report ‘Artificial Intelligence: The new frontier for hedge funds’ in which we had the opportunity to discuss your unique artificial intelligence (AI) strategy in detail. How has the reception been since then to your strategy? A year later, are investors more comfortable in talking about AI hedge funds?

    My strategy is based on the idea of how the market price moves in general. Many investors keep eyes on market price movement. When market price moves, some investors start to react to it and trade some amount, which causes the next movement of the price and other investors become involved in it. Therefore, there is a feedback system which makes all the investors interact with each other throughout the market price movement. A trend/momentum generates when the feedback is positive and mean reversion generates when it is negative. Generally speaking, investors tend to understand momentum and mean reversion as different objects but I believe they are just the different phases of the same origin. This thought is described in a mathematical equation and AI/machine learning methodology is applied to lend it predictive capability even when the market environment is changed.

    As to a pure AI strategy, many people have been interested in my strategy and I have received many inquiries globally since we last discussed the nuts and bolts of my strategy in the Eurekahedge report. I think investors have become more comfortable to talk about AI hedge funds. ‘AI’ has become a buzz word and any one can try their hands on an AI program because there are plenty of open source codes available on the Web. So, I believe utilising AI/machine learning in the hedge fund space is almost the need of the time.

  2. Hedge funds across the board delivered strong returns in 2017 as markets continued their uptrend and the year ended without much incident. How did 2017 turn out for your strategy with regards to both the momentum and mean-reversion components of your AI strategy? What went right and what went wrong?

    My strategy requires volatility in the market whereas the market in 2017 is known as one of the historically lowest volatility environments. For example, the volatility of the intraday return in 2017, the return from open price to close price of the TOPIX, was specifically lowest at least in these two decades, which is obviously an extraordinary situation. Even if there were imminent risks of ballistic missiles and nuclear threats, the investors easily became indifferent to them. In the machine learning point of view, this super low volatility phenomenon causes three major problems. Firstly, smaller amplitude of mean reversion cycle makes the information in the input time series less meaningful. Secondly, too detailed optimisation for shorter cycles tends to underestimate longer scale movement which results in a larger loss than all gains from shorter cycles, i.e. Detailed Optimisation Problem (DOP). Thirdly, the noise of the market such as the effects of idiosyncratic events, random political speaks, etc. becomes prevailing in the input time series, which makes it difficult to extract embedded meaningful information from the time series.

  3. Could you walk our readers through the key enhancements to your AI model over the past year and what necessitated them?

    How can we deal with the low volatility environment without causing DOP, detailed optimisation problem? This was the key question in 2017. In other words, local optimisation is too dangerous to dig into but global optimisation might be regarded as useless from client investors because it takes too long time to recognize its efficacy. In order to deal with it, the model requires to evaluate microscopic, mesoscopic and macroscopic movement evenly at the same time. In mathematical translation, you need to find some mathematical form describing the governing dynamics which is invariant in scaling of the scope so that you can dynamically renormalize the scale to avoid local optimisation. I know this concept is very difficult but I believe this approach tells the truth in the ultimate.

  4. So how have these upgrades enhanced the quality of the trading signal generated and do the numbers support this?

    I developed a model taking account of the requirement above last year. As to trading signals, the model enhanced its power of expression to catch the variety of possible scenarios which might happen in the nearest future. After the implementation of the enhanced model at the end of October 2017, the model seems to be able to deal with the low volatility environment and continues to deliver positive monthly returns so far.

  5. Your fund currently specialises in Japanese markets, i.e. Topix futures, yet the fundamental premise of your investment strategy, i.e. mean-reversion and momentum characterize every market. Could you share with us your reasoning behind the decision to limit trading to just the Topix futures; and if the model can be applied across other regional markets as well?

    The TOPIX consists of over 2,000 names and represents Japanese stock market. I thought focusing on TOPIX futures would be a good demonstration for the first new methodology in a simple manner. I think the model also can be applied to other instruments and other major equity futures too. For example, a simulation of same AI strategy investing in Asian equity futures such as MSCI Singapore, Hong Kong Hang Seng, MSCI Taiwan and Nifty50 shows strong risk adjusted return over time and the simulation is conducted by “Walk Forward Test”1 which demonstrates the model’s adaptive capability to market environmental change to maintain predictive capability.

  6. Global markets have receded quite sharply in February with a spike in volatility and a pullback in major equity market indices across the board. How has the Simplex Equity Futures fund performed under such volatile conditions especially with regards to the trading signals generated by your AI hedge fund on 2 February 2018 and 5 February 2018?

    The positive momentum monitored by our model had been decreasing since 25 January 2018 and at the end of 2 February 2018 (Friday), the fund held 50% of long exposure and the exposure was all squared at the opening of 5 February 2018 (Monday) and it kept the zero position all day long so the major fall of TOPIX futures (-4.6%) from the opening of 5 February 2018 towards the opening of 6 February 2018 was completely quarantined from the fund. On the opening of 7 February 2018 when we made 50% of short exposure because our mean reversion index reacted, the market opened 3.0% higher than the last closing price and fell -2.0% toward the closing of the day when we covered major short exposure.

  7. Please share with our readers how your AI strategy incorporates risk management into the investment process to guard against drawdowns. How successful do you think it has been?

    This strategy does not have fixed limit such as -5% loss cut rule. The exposure to be taken is based on the prediction of short term directionality of the market and the real loss cut happens before it accumulates significant losses. Despite this, continuous losses can happen but the magnitude of draw down was less than the range originally estimated by the Walk Forward Test. So, I think it is successful thus far.

  8. What kind of investors does your fund cater to and what has been the reception so far with regards to capital raising? What are some of the challenges you face when marketing an AI hedge fund to investors?

    As long as the market condition is favourable for index investors just like the market in 2017, investors have less incentive to invest in newer hedge fund strategies. Our clients are Japanese financial institutions whose investment process is strictly well organised therefore explanation of the strategy is challenging. Although my program is not ‘a black box’ just like normal neural network, explanation of the model is not so easy as conventional quant models because the basic concept come from both physics and machine learning.

  9. What are your thoughts on the market outlook going into 2018 and how well is your strategy equipped to handle any turbulence along the way?

    After we experienced the plunge in February, an easy recovery to the situation in 2018 with super low volatility appears to be difficult. So, I hope a humble up trend with larger volatility toward the end of 2018.

  10. Lastly, some would argue that true AI applications to financial markets are yet to arrive, with most of the recent developments being instances of ‘IA - intelligence automation’ as opposed to ‘AI – artificial intelligence’. Moreover, others doubt if the machine can harness the kind of finesse that humans bring to the art of investment. What are your thoughts on this?

    This is an interesting argument. From a programmer’s point of view, I don’t see big difference between AI and IA. If you know how it works, it seems to be ‘IA’. If you don’t, it seems to be ‘AI – autonomous intelligence’ just like your own brain which is a bio machine though. However, as far as investment is concerned, a machine can provide high speed, accuracy and stable results that humans cannot achieve. You at least need to harness the finesse of machines, otherwise you will lag behind your competitors sooner or later.

    As to the future of AI investment strategy, machines will become much more sophisticated than imagined today. For example, someone may say that many quant strategies suffer from crowding problem so the AI strategy will have the same problem down the road. I do not think so. The reason some investment strategies get crowded is because humans crowd in the strategy. The low volatility market in 2017 followed by the big plunge in February 2018 was set up by typical human crowding, in other words human cognitive bias called ‘Overconfidence and Biased Secondhand Knowledge’.2 Human investment managers tend to follow major/popular investment strategies because it is easy to explain it to others as an accountable/fiducial investor except for declaring and implementing minor different ideas. On the other hand, AI can be free from those human cognitive biases and it will be able to take account of many other possible scenarios which are sometimes difficult to explain to others in a simple way because the reality of market is not that simple at all. Therefore, the behaviour of AIs will be very different from each other and if there is a flash crash in a few milliseconds, some AI might find it as a good trading opportunity and the price might come back again in a very next millisecond by a counter flash, faster than your eyes blink!

    I have a little concern in terms of the ‘explanation’ of AI investment strategy. Current global media might report bots as villains because it is easy to blame unknown objects as cause of catastrophe but current bots are just designed to take advantage of unduly crowded humans. A simple-minded crowding of humans might be more dangerous just like CDS bubble followed by Lehman shock when CTAs made a fortune and the Eurekahedge CTA/Managed Futures Hedge Fund Index3 gained 19.39%! If regulators find the value of diversification of strategies rather than simply their acceptability and understandability, AI will be able to provide the value to the market in the future!

Interview with Atsuo Ogaki and Damien Loh, founders at Ensemble Capital Pte Ltd.

Ensemble Capital Pte Ltd. is an independent hedge fund and investment management firm in Singapore founded in 2017.

  1. Please share with our readers a bit of background on the Ensemble Capital Fund, the fund’s key personnel and your decision to launch an AI/machine learning hedge fund strategy run out of Singapore.

    Atsuo Ogaki, CEO of Ensemble Capital Pte. Ltd., started his 29 year trading career in 1988 in Tokyo at JP Morgan (formerly Chase Manhattan Bank) where he spent 23 years trading FX Options (specialising in Japan macro markets) and running the team in Asia. He spent a considerable amount of time outside of Tokyo as well, with over six years in locations such as Singapore, London and New York. He has a track record of being a consistent revenue producer and his most profitable years have been during uncertain times. His last role was at Nomura where he was the head of FX Japan, managing a team of 30 people. He was also a representative in Japan among the Global FX management committee.

    Damien Loh, CIO of Ensemble Capital, graduated from Cornell University with a degree in Computer Science. He has traded FX options for JP Morgan for 15 years in New York, Tokyo, London and Singapore. During his tenure he traded both emerging and G20 FX volatility products across the spectrum from vanilla options to exotic products. Over and above his responsibility of being a market-maker, he was instrumental in developing the algorithms for offering FX options on the bank’s electronic platform. He has a track record of being a consistent revenue producer (no negative P&L years) and his most profitable years have been during uncertain times.

    The development of sophisticated techniques in AI, increases in the computational powers of GPUs, and the availability of data have all converged in recent years to make it an opportune time to launch an AI-based fund. Ogaki and Damien noticed that banks and macro hedge funds have been slow in the uptake of this technology which was the motivation to found Ensemble Capital Pte. Ltd., a registered fund management company based in Singapore. Ensemble Capital manages the Ensemble Violin Fund, an absolute return global AI fund. Our background in finance and formal training as computer scientists put us in a unique position to be able to communicate with our deep learning scientists to customize AI techniques for trading.

    In keeping with its goal of looking to be an intelligent island, Singapore has attracted top AI talent from around the world and this was one major reason in choosing Singapore as our headquarters.

  2. Please walk our readers through your investment strategy and the type of instruments and markets you focus on. How many trades do you make on average and what kind of leverage is utilised in the strategy?

    We have an ensemble of models for each currency pair which together come up with a forecast. The forecast comprises of a signal for direction, magnitude, conviction, and time horizon. Trades are expressed mostly through option structures which vary depending on the events coming up, level of implied volatility and axes in the markets among other factors.

    Given the founders’ background, the Fund currently focuses on currency and currency options. In time, as the investment process is vindicated by a good track record, we will look to branch out to other macro instruments as the deep learning models would be just as viable in other markets.

    We do not have fixed leverage limits but we have traditionally used about 2-3 times notional leverage for our backtest and have been adhering to that range since our start.

  3. Leading from the above, could you tell us more in particular about the AI component in your strategy and how it has the potential to enhance your returns?

    Our strategy uses an ensemble of models as opposed to a single model. This approach has a few advantages:

    • We use different model architectures that best suit the different data modalities. For example, if the data is in the form of a time series, a variant of a recurrent neural network would be best suited for this task.
    • Different data inputs or features for each model allow it to come up with the forecast from a different perspective. An analogy would be having a value investor, momentum investor, flow trader and economist come together to make a composite forecast.
    • Component signals from each model can be adjusted based on performance to improve the accuracy of the composite signal.

    The AI aspect allows us to have an informational, behavioural and an analytical edge:

    • Informational: AI gives us a framework to use the wealth of information from disparate sources (sentiment, price, momentum, seasonality) to answer a singular question, namely what the expected level of a currency pair is for a given time frame.
    • Analytical: The specific use of deep learning within AI allows the model to recognise subtle relationships between the features and each layer builds upon the previous one to go from basic to more abstract concepts.
    • Behavioural: The forecasts instils a discipline in our trading to ensure are we using a consistent investment process to achieve our returns

    Building our AI models in-house enable us to quickly retrain our models to respond to any regime change in the markets.

  4. The strategy appears to take a flexible approach in terms of keeping human oversight over trading signals generating through deep learning/AI applications. How does this provide edge to your strategy in comparison to traditional quant hedge funds utilising big data and those on the other end of the spectrum utilising AI/machine learning methodologies in their trading processes?

    We emphasise that our portfolio managers would not go against the model forecasts and that the flexibility comes predominantly from how the trade is expressed. This ensures that man and machine are working in concert as opposed to countervailing each other. Human oversight is needed to adjust the level of capital deployed when the main driver of markets is a phenomenon that the model had not accounted for. For example, if a natural disaster has occurred or a one-off referendum has been scheduled, these are events that are very challenging if not impossible to factor into any model.

  5. Back tests of your strategy since 2007 indicate healthy performance in terms of annualised returns and strong downside protection. E.g. simulated returns from 2007 to 2017 indicate annualised returns of 11.92% with high Sharpe and Sortino ratios (1.95 and 3.27 respectively) with a maximum drawdown of 6.10%. Could you share with our readers how your experience has been since the fund first starting trading, especially given all the volatility we have seen in recent days?

    At this time of writing, it has been only two weeks of live trading and we are still incrementally increasing our capital allocation. Hence, I do not think any performance would be representative. What is evident even in this short period however is that the framework has worked as expected and that our modestly positive returns so far is in keeping with the uncorrelated returns.

  6. One thing that strikes out when viewing your simulated returns is how annual returns since 2012 have trended lower into low-to-high single digit numbers, while the period before that (2007 to 2011) has generally produced spectacular double digit returns. A quick look on my end across the Eurekahedge Global Hedge Fund Database suggests this has been the case for pretty much every hedge fund out there. To what would you attribute this return disparity post 2012, especially in the context of your strategy?

    We attribute the large returns in that period to dislocations in the market during volatile moves. The resultant inefficiency in pricing makes it a very conducive environment to trade options with its inherent asymmetric pay-out. This also fits nicely with our aim to provide uncorrelated returns compared to traditional asset classes. Recent returns have been more modest due a confluence of two factors, specifically, efficient market pricing due to positive risk sentiment and markets moving more on sentiment rather than factors. This has been identified by us and our next addition to our ensemble will be more sentiment focused. The ensemble framework allows for easy additions of new models using different information and architectures.

  7. What sort of risk management tools and practices have you implemented to safeguard your portfolio returns? In particular, what processes do you have in place to monitor the robustness of trading signals generated through the AI models, and how often do these models undergo any enhancements if necessary?

    We learn on in-sample/training data and tested on that data as well as an out-of-sample/test set. Part of our selection process is to ensure that the drop off in performance between the training and test set is minimal which helps to verify that there is minimal overfitting and model has generalised patterns instead of merely ’memorising’ the training set.

    Because we build our AI models in-house, we can make minor and major adjustments to our models quickly; each day we look at how the each of the underlying models in the ensemble of models has done relative to its previous forecasts and adjust the weights assigned to each of them accordingly based on their recent performance. At regular intervals through the year we also look to re-train the models with more updated data to ensure it is learning the most recent behaviours in the market. During shifts in the market regime we have the flexibility to do an unscheduled retraining of the model and could look back to periods in time in the past where markets are behaving similarly to current markets for training data. . For example, if bonds and equities are moving in a negatively correlated fashion we could look for periods in the past where the markets behaved similarly. Note this is a highly simplistic example and we have a much richer metric of what constitutes similarity between markets than just one dimension.

    Each forecast/inference comes with it a conviction score; a low score means that the forecast is relying on a small subset of features to make its inference whilst a high score means a broad set of the features agree with the inference. This allows us to decide on the amount of capital we would like to allocate on the trade.

    On the traditional side of things, we watch stop losses on individual trades and have drawdown limits on overall portfolio basis to limit losses. We also monitor concentration risks using option greeks and VAR numbers as metrics.

  8. AI hedge funds appear to offer uncorrelated returns not just to traditional markets but to their predecessor traditional quant hedge funds as well which adds to their appeal for investors. How has your experience been with regards to raising capital for the fund and what is your planned capacity for this strategy?

    Investors have been very receptive to our approach and we are confident in being able to fill up our inception class which offers lower fees for early investors who support us from the very beginning. Macro markets offer depth so we don’t see capacity constraints until the fund becomes has more than a few billion in AUM especially given our plan to branch out of currencies.

    We believe that part of the appeal is our emphasis on volatility products which provide a viable uncorrelated alternative for investors to diversify their portfolio returns. Historically, equity and bond returns used to be uncorrelated. However, monetary stimulus over the past few years have linked all traditional asset classes together. Investors are looking for alternative asset classes to invest in to diversify their portfolios.

  9. It seems that unlike other AI hedge funds which tend to minimise human input in their trading process, Ensemble Capital maintains a fair degree of scepticism with regards to the amount of responsibility that can be outsourced to the AI. Could you share on the challenges you have encountered so far with regards to applications of machine learning and artificial intelligence techniques in trading?

    We do not think our opinion contradicts other AI hedge funds' approach. Deep learning works best when a simple question is posed to the model to reduce the number of dimensions it needs to search for an answer in. If we had framed the question as “what is the best option structure to trade in across all the currency pairs across all tenors?” we would expect to see very poor forecasting results. We believe humans still excel at rapidly taking into account new phenomenon even when available information is minimal.

  10. Going forward, what are your expectations for your strategy in 2018 especially given the recent concerns over liquidity in the volatility space? And from a more long term perspective, how do you expect the power balance between man and machine to evolve as AI strategies continue their move into the mainstream?

    First and foremost, we are not a high frequency trading fund and we do not rely on short-term liquidity of the asset class to express our views.

    That said, the liquidity in currency options is deep enough for us to reach AUM in the billions before posing an issue. Trades can also be expressed in the underlying directly as well in extenuating circumstances.

    We think this issue is too often framed in an ‘us vs. them’ perspective when it is more likely we will see man and machine working in a collaborative manner to be able to be more efficient and productive. Whist we might cede some responsibilities to the machines as they become more sophisticated, it in turn frees us up to use the technology to work on interesting issues.

The full article inclusive of all charts and tables is available in The Eurekahedge Report accessible to paying subscribers only.

Subscribers may continue to login as usual to download the full report and non-subscribers may email to enquire on how to obtain the full research report.

2Evidence-Based Technical Analysis, David R Aronson, 2007, Pan Rolling