Forward-looking market participants are using artificial intelligence to streamline the workflow of the trading lifecycle, and to assist human traders in deploying algorithmic trading strategies.
Portware, a provider of trading technology, has brought out a commercial version of this idea in the form of Alpha Vision, an algorithmic optimization solution that uses predictive analytics to automatically select and implement the optimal execution strategy.
“Over the past 12 months, as we’ve looked at what would drive the next wave of innovation in trading, it became evident that traditional execution management systems (EMS)—including those from Portware—have had a profound impact on buy-side trading desks by providing greater efficiency, reducing operational risk, automating workflow and providing access to global liquidity providers,” said Harrell Smith, head of product strategy at Portware.
“Yet, while there have been tremendous benefits to both the buy side and sell side through the evolution of EMS technology, it has taken workflow automation as far as it could,” he said.
The industry is in the midst of an operational overhaul designed to significantly boost front office productivity.
“With fewer people available to navigate through the mountains of data and systems, front offices are in demand for intelligent, signal-rich infrastructure where machine learning-based systems provide a deep bench for research and analysis in support of a thinly staffed front line,” said Jon DiGiambattista, capital markets expert at Opera Solutions, whose Global Markets Signal Hub uses machine-learning science to identify signals in noisy time-series data.
Right Strategy at Right Time
Machine learning is playing an ever-increasing role in financial markets, albeit in ways that are different from what was envisioned a decade ago.
“Machine learning is very adept at identifying patterns, so where we are seeing the most beneficial application is where time series can be observed across an exhaustive set of data in order to distinguish behavioral anomalies among individual assets (e.g. a specific bond issue), asset classes (e.g. mortgage backed securities) and macro factors (e.g. housing market),” said DiGiambattista.
Although the sell side has invested heavily in developing sophisticated and customizable algorithmic strategies, the responsibility falls to the buy-side trader for selecting the appropriate algorithm for every order.
“The trader is responsible for monitoring every order and ensuring that an algorithmic strategy continues to be appropriate given changing market conditions, changes in liquidity profiles, price and volume changes,” said Smith at Portware.
By dynamically switching between algorithms and adjusting to changing market conditions in real time, Portware’s Alpha Vision maximizes alpha capture while minimizing impact costs, adverse selection, information leakage and the impact of high-frequency trading.
“In today’s high-speed electronic markets, it’s impossible for a human trader to make sure that every algorithm that they select is correct for the given set of market conditions,” said Smith. “Alpha Vision takes trading to the next level by introducing true artificial intelligence into the trade workflow for the first time.”
There can be a variety of automation scenarios. One is to have a “classifier” that scans the market looking for an opportunity to go short or to go long.
“The described principle can be conceptualized as a problem of pattern recognition: a significant amount of data is available from the market; some data is associated with specific events, e.g. change in the price of an asset, and can be used in trading,” said Vadim Mazalov, research and development specialist at financial technology firm Cyborg Trading Systems, and PhD Student in computer science specializing in machine learning at Western University in London, Ontario.
Filtering Out the Noise
The model should be able to compare the incoming data with the training data sets and make a decision in real time.
“The classifier may predict that the prices of an asset will go up or will go down in a certain time interval,” said Mazalov. “Short-term forecasts, as opposed to the long term, are more feasible to achieve: the amount of contributing factors is smaller and their values are easier to measure.”
Also, the possibility of a negative impact by outside factors, e.g. an unfortunate news release, is “asymptotically lower and has less effect in a short term”, said Mazalov.
“In our experiments, features selected from a stock can be sufficient to predict the change in price of the stock in about a second in advance, which is valuable in a high-frequency setting,” he said. “To increase the prediction interval or predict beyond a single price change, one needs to include into consideration correlated equities.”
Neural networking, a form of artificial intelligence that mimics the workings of the human brain, is also being employed in automating the trading lifecycle.
For example, TradingSolutions, a product of software provider NeuroDimension, combines technical analysis with technologies using neural networks and genetic algorithms to learn patterns from historical data and optimize system parameters.
The system works with stocks, futures, currencies and other financial instruments. It can also build systems for U.S. and international markets.
“We’ve combined the trading signal generation capabilities of TradingSolutions with our automated trading platform, such that the entire trading process can be fully automated to remove much of the discretion and emotion from one’s trading,” said Gary Lynn, chief executive of NeuroDimension. “Of course, there is always the decision process involved of choosing the right mix of assets and trading models and adjusting the mix according to the actual performance.”
The biggest pitfall is over-fitting the model to past data.
“This is why it is important to keep the number of weights (i.e., parameters) of the neural network to a minimum and to leave a significant portion of the data out for out-of-sample testing purposes,” said Lynn.