A BRIEF HISTORY OF… ALGOS

Execution algos

Electronic trading started to mass-develop in the ‘80s across asset-classes, and accounts today for a vastly dominant portion of the overall trading activity. As trading volumes increased and more electronic markets progressively opened, investors requested improved execution capabilities by controlling execution cost and market risk, and the business of execution algos took off in the early ‘90s.

The main consequence overall was the compression of execution margins, as high touch voice services were progressively replaced by electronic execution and as the offering in term of algos widened. Today, most major banks as well as online trading platforms propose to their client a full suite of execution algos, across all electronic asset-classes, i.e. equities, forex and more recently bonds. This has become an extremely competitive business, as illustrated by the falling average cost charged on trading through algos.

The main categories of execution algos are transaction cost reduction and liquidity-seeking engines, whereby large orders are broken down into small orders and placed in the market over time, and allocation is optimized across available trading venues (standard TWAP, VWAP, POV, and more advanced developed by banks like DB’s Stealth, CS’s Sniper, etc…)

High Frequency Trading algos

At the same time, many institutions (banks, funds) saw new profit opportunities from deploying algo strategies not as a facilitation service, but as alpha generating tools: mean reversion, index arbitrage, statistical arbitrage… While these strategies were implemented in a rather simple way initially, the competition fuelled the deployment of ever more complex algorithms and a race for speed (reduced latency) and optimization of execution (smart routing), in what is now called High Frequency Trading (HFT).

Main elements taken into accounts by HFT strategies are:

  • Order book dynamics: number of orders on each side, size of the orders, as well as the cancellation dynamics of orders, time the orders are active, time before orders are crossed, etc… can contain predictive information with regards to securities price movements and algos scan across all these dimensions for price discovery purpose.
  • Trade dynamics: similarly, trade patterns can contain information about future price movements. For instance, the origination of most trades (buyers or sellers), the size of past trades and time between them, the type of trade (limit, market, stop order), all can contribute to information about future price
  • Past stock returns: past stock returns, in combination with past order book dynamic and past trade dynamic, can be used to anticipate future price moves, and decide whether momentum or mean-reversal type of strategies are more likely to perform at a given
  • Cross stock correlation and cross asset correlation: for two or more assets with high short-term correlations, high frequency traders seek trading opportunities by trying to be the first to react to price movement on one asset or on one market, before it impacts correlated assets or other markets.
  • Cross exchange information delays: with market fragmentation and multiplication of alternative trading venues, high frequency traders seek to exploit speed of information differentials to uncover profitable

Empirical evidence suggests that HFT contributed to reduce bid-offer spreads and is a substantial source of additional liquidity. However, this added liquidity can prove volatile and actually disappear when most needed, in time of market turbulence. Also, there is a growing sense that HFT can at times represent a threat to markets, as illustrated by the US flash crash, or the fake White House bomb report incident.

Recently, two additional types of concern arose that darken the future of HFT as it exists now.

·       Technical

As the HFT industry matured and becomes more and more competitive, the ability to further compress latency is reaching physical limits, meaning innovation and differentiation driven by the sole speed of information / execution has reached a plateau. At the same time, the number of participants has grown dramatically, resulting in the overall decrease of profits directly linked to HFT.

– In a move to address concern about fair treatments of all participants, some of the biggest FX platforms (EBS, Thomson Reuters) have recently introduced latency floor and MQL (minimum quote life) constraints. Latest academic research shows that this last rule caused the actual trading volumes on the platform to decline, suggesting some market participants moved to other venues without MQL

– Similarly, some exchanges, like in Italy, introduced constraints around the order cancellation ratio, whereby market participants who cancel a lot of orders prior to execution are financially penalized.

As these initiatives are likely to spread, HFT as it existed until now will be challenged.

·       Regulatory

HFT has also recently been the cause of regulatory concern, as regulators worldwide question the integrity of an activity which until now has remained largely unregulated. At the European level, MiFID2 directive has a requirement that algorithmic traders should report to the relevant authorities about their strategies, limits or trading parameters. Similarly, the FCA recently announced that algorithmic and high frequency traders will now be part of its certification regime, which requires banks to assess and then certify whether individuals are suitable for particular rule, and assessments are updated annually. Regulators around the world, including in the US, Canada, Japan, Australia, France are in the process of reviewing regulations around HFT, and some are already taking measures for better control.

In parallel to the explosion of HFT volumes and related strategies that leads to overcrowding symptoms, recent technological innovation is spurring the emergence of new types of algos: big data and artificial intelligence developments bring a new future for trading algos.

Artificial Intelligence trading algos

·       Concept

Traditional quantitative trading is constructed around a mathematical model, designed and developed by human beings, that helps identify trading opportunities. As and when market environment changes, the model needs manual updating and optimizing to new conditions. In the case of artificial intelligence trading programs though, the initial code is developed by human-beings, but it is changing and improving itself over time, with positive feedback loops based on self- awareness and self-adaptation.

·       Data used

With the advent of big data, input to algorithm is no longer limited to structured data (such as market and trade data, companies’ financial information etc…). All sort of data (news, sentiments, information), under any format (picture, video, database…) can be stored in unstructured databases and used as input for algorithms. The volume of information as well as the time to screen through these huge amounts of data are no longer a limiting factor, thanks to technologies like massively parallel processors, in-memory databases or NoSQL.

·       Models

Many algorithms used in AI algo trading have been created long time ago, but big data, by making available massive amounts of data cheaply and quickly, is allowing them to live up to their full potential. Most common techniques used are fuzzy logic systems, decision trees, induction rule sets, neural networks, Bayesian networks and deep learning algorithms, which allow to spot trends, patterns, and more generally speaking extract salient features from the data which have predictive power with respect to the considered outputs.

·       Self-awareness and self-improvement

This is a key part of the technology. The initial algorithm, which is an optimizer under constraints, is designed to maximize profits while controlling for a set of risk measures (drawdown, volatility, diversification…). It is calibrated with available information at that time. However, as market conditions change and new information is available, the logic or neural networks previously employed might not work any longer. Artificial intelligence feeds on itself, in the sense that information about the model behaviour and performance is used to improve the model itself.

·       Challenges

The combination of IT infrastructure, machine learning, mathematics and quantitative finance skills required for these projects make it difficult for someone to just come up with a machine learning system. Also, as AI algos constitute a high stake subject for most involved players (large R&D investments, high expected returns relying on developed techniques to remain undisclosed), the industry is very secretive about the latest developments in the field.

·       Current disclosed projects

Some hedge funds are investing heavily in the technology and a few projects are publicly reported.

  • Bridgewater Associates: the world’s largest hedge fund launched a six-strong AI unit led by David Ferrucci, former IBM’s Watson team
  • Renaissance Technologies: the fund’s co-CEOs Bob Mercer and Peter Brown developed language- recognition programmes at IBM before they joined Renaissance, and the firm is reported active in the
  • Two Sigma: one of its co-founders, David Siegel, has a PhD in computer science from Massachusetts Institute of Technology’s AI laboratory, and the firm runs some money based on
  • Highbridge Capital Management: the fund is known to be working with San Francisco-based Sentient Technologies to develop investing strategies using

Algo Trading related fintechs

Below are what we think are the most interesting startups involved in algos at the moment, from statistical and data mining algos to advanced machine learning driven algos:

 

  • Scalable Capital’s proprietary technology dynamically allocates each investor’s portfolio based on a quantitative measure of their risk appetite. The technology uses forward-looking projections, based on recent market developments, to measure the level of risk in the ETF products the client is invested in, and then reallocates their portfolio according to their risk limit. In contrast to traditional wealth managers, Scalable Capital adopts a fluid approach to the weighting of asset classes in its portfolios. This allows investors to capitalise on markets where risk is rewarded, and limit exposure to excess risk in more volatile conditions. technology is an institutional class investment product, available, for the first time, to retail investors, at a fraction of the cost. Scalable Capital received regulatory approval from the FCA in February

 

  • Darwinex: is a trading platform that matches traders and capital providers. It provides markets on fx, equity indices and commodities, and designed a unique trading strategies scoring algorithm that allows investors to allocate to the best managers based on their specific risk appetite

 

  • WooTrader is a system that uses predictive analytics to model the stock market and rank individual stocks by current and expected performance. The system uses hundreds of individual screeners based on fundamental and technical analysis, guru strategies, options volatility and ratios, social sentiment for news articles, blogs, twitter, and stock twits; and many more areas to analyse each equity and then assigns a weight to each area based on what is currently driving the market before grading and ranking the stock. This allows an individual analyst or advisor to scan thousands of equities in a matter of minutes and instantly see what forces are currently driving the markets, which sectors are performing best or underperforming, and even generate pdf reports based on their research or portfolio positions. Woo Trader also integrates with brokers to allow actual portfolio management right from the dashboard.

 

  • Quantopian is a crowd-sourced quantitative hedge fund. They provide capital, data, and infrastructure for quants to research, code, test, and trade algorithmic investing strategies. They have an engaged community of 50,000 quants who discuss concepts and practice, and learn from peers and experts. They provide capital for their community to trade with, both through a monthly contest and through an allocation process for high-quality

 

  • Quantconnect gives investors access to algorithmic trading. They serve financial engineers (“quants”) and investors with a platform to design algorithms and investors to fund

 

  • Inovance helps developing new trading strategies through data mining and machine learning algorithms (both supervised and unsupervised learning).

 

  • Symetrics: provides cutting-edge behavioural algorithms that feeds econometric models to assist portfolio manager analysing and hedging current and future risk

 

  • Walnut Algorithms is a technology firm focused on applying the latest advances in data science and machine learning research to the financial markets. Walnut Algorithms develops sophisticated trading models able to scale over numerous assets globally. The strategies are designed to spot intraday patterns forming in the financial markets with high levels of confidence. The large amounts of intraday data points available allow them to train robust machine learning models and avoid

 

  • BondIT is the algo-advice solution for bond portfolios managers and advisors. BondIT empowers fixed income investment managers with data-driven portfolio construction, optimization, re-balance, analysis and monitoring capabilities. The BondIT solution is designed to enhance the entire bond portfolio management life cycle with actionable analytics and algorithmic recommendations for buying, selling and replacing bonds in the portfolio. It also provides on-the-spot comprehensive analysis and improvement suggestions for imported and pre-existing portfolios. The solution is based on advanced proprietary machine-learning algorithms developed by a team of financial and data analytics experts.