The Artificial Intelligence (AI) and Machine Learning (ML) is quietly revolutionizing nearly all areas of our lives. Do you know latest trading algorithms are using it extensively?

Machine Learning hedge funds already vastly outperform generalized hedge funds as well as traditional quant funds according to a report by ValueWalk.

While it may be risky for hedge funds managers to outsource investment and risk assessment decisions to robots - AI systems can serve as an invaluable help for humans during the decision-making process.

This article is intended to be a business overview of what Machine Learning is capable of doing in market trading with a high-level technical explanation.

AI/Machine Learning Hedge Fund Index vs. quants and traditional hedge funds. Source: Eurekahedge

Below are the most important AI applications.


A system can monitor stock prices in real time and predict them based on the news stream (AZFinText). The experiment used Support Vector Machine (SVM) to trade S&P-500 and yielded results that cannot be neglected - below is the table presenting how it performed compared to the top 10 quantitative mutual funds in the world:

Experimental trading strategy using Google trends

There are a plethora of articles on the usage of Google Trends as a sentiment indicator of a market.

For example, the experiment in this paper tracked changes in search volume of a set of 98 search terms (some of them related to the stock market). Term "Debt" turned out to be the strongest, most reliable indicator and market operations were performed using DJIA.

Below is a cumulative performance chart. Red line depicts "buy and hold" strategy. Google Trends strategy (blue line) would have yielded a profit of 326%.

Performance of Google Trends strategy

Advent of robo-advisory in wealth management

AI-powered portfolio management and online advisory systems will play a significant role in helping people with relatively small savings (less than $250k) who cannot currently afford professional consulting services. By bulking together small accounts which, on their own, are unprofitable, an investment firm can collect sizable additional cash influx.

This management could potentially undermine to some degree the position of traditional wealth advisers and lead to personnel reductions, yet the more experienced advisors have nothing to worry. Human input is highly desired in crucial life decisions like investments. AI portfolio management systems will coexist with professionals and be used for a subset of decisions, e.g. less risky investments.

To give an example, in the US, Wealthfront added an extra facility that harvests tax losses that it says was previously only available to portfolios of $5 million and above, claiming it can add extra 1.55% a year to portfolio return.

The vast majority of (online) advisory programs are just algorithmic software or worse - passive portfolios - they certainly fill the need in the market, but are relatively blunt instruments. Deep Learning (DL) being on the forefront of AI, is predicted to constitute a core wealth management and planning platforms of the future.

I'm going to learn ML myself.

Well, of course, you can. There is a popular Udacity course on how to apply the basis of Machine Learning to market trading. To understand more from these materials, as a starting point we recommend:

Machine Learning for trading is a vast topic that takes the time to master, and these are just introductory materials to get started. If you want to speed it up, you can hire a consultant. Remember, to ask tough questions before starting a project especially because your money is on the line.

Or, just schedule a short call with us to explore what can be done.

How is Deep Learning different and why I should care?

Deep Learning is on the forefront, on the bleeding edge of technology.
In contrast to a standard algorithmic trading or Machine Learning, Deep Learning-driven system does not need specific rules, indicators or features (feature engineering process). It's able to invent them itself saving analysts huge amount of time and, it performs well when the scale of the data increases (often, using DL requires a larger volume of data comparing to classical ML methods).

DL system is an Artificial Neural Network with many (often hundreds) of hidden layers. After it's constructed, it needs to be trained.

Training is a process of feeding the network with data and right answers (in this case - historical market data), and "penalizing" it for every incorrect prediction in a process called back-propagation. It often requires significant GPU unit(s) (for heavy computation), but nowadays they are accessible on the Cloud (e.g. Amazon P2). The outcome of the training is a computed matrix of weights defining the connections in the network.

See a short introduction on how to train a Tensorflow network on Amazon P2

What is more, more often than not, Deep Learning provides an End-To-End solution for a given business/engineering problem whereas using pure Machine Learning may require breaking the pipeline into more, distinct steps and phases.

The image below presents a simple artificial neural network with two hidden layers:

neural network

... while the next illustration presents a GoogLeNet - a network with over 100 layers that was the winner of ILSVRC 2014 (ImageNet Large Scale Visual Recognition Challenge). It's successor - Microsoft ResNet - outperforms human intelligence in object recognition and visual tasks.


Another interesting difference is interpretability. Deep Learning methods - despite often supreme in performance to ML - do not give a clear explanation or reason - why an accurate prediction is generated (and what particular neurons are modeling) - so it's hard to interpret the results. Whereas for Machine Learning models - using e.g. linear/logistic regression or decision trees - generate predictions that could be easily interpreted by researchers - therefore they are still well represented in the industry.

I need more specific examples applicable in my industry


This paper describes how Deep Neural Networks (DNN) were used to predict 43 different Commodity and FX future mid-prices.

Another experiment describes trading on Istanbul Stock Exchange with NN and Support Vector Machine (SVM).

Interestingly enough, this paper presents how genetic algorithms support vector machine (GASVM) was used to predict market movements.


Implementing a successful Machine Learning investment strategy is hard. You will need extraordinary talented people with experience in trading and Data Science. By using it, your portfolio can yield a greater alpha.