Deep Learning Algorithms: The Future of Financial Investment?

This article was written by Harry Chiang, a financial analyst at I Know First.

Deep Learning & Finance

Summary:

  • What is Deep Learning?
  • Deep Learning Applications Today
  • Deep Learning & Financial Markets
  • I Know First’s Role in the Machine-Learning World

Deep Learning Algorithm

Picture an Olympic athlete running along a track. Most humans would understand, perhaps even intuitively, that when he or she runs there is a certain path of movement and way in which he or she is interacting with the environment. The athlete will follow the curvature of the track. The rules of the physical world bind the athlete. They dictate how his or her body moves and how his or her feet must move along the rubber. A machine, however, would struggle to understand all these small details. Machines are very good at running many complex calculations. However, they struggle when it comes to predictive patterns and interactions.

Thus, it is understandably impressive that in 2016, MIT CSAIL researchers managed to develop a deep-learning machine that can predict the future. Perhaps not exactly predict the future, but given a still image from a scene, it can create a brief video that simulates the next action in that scene. For example, a train moving along a track, or waves on a beach. These are obvious intuitions for us humans, but machines struggle with computing this. The researchers trained the algorithm using 2 million videos, a year’s worth of footage. When tested, human subjects deemed the generated videos to be realistic 20% more often than a baseline model.

This predictive model is important for future AI/deep-learning research because it will allow researchers to scale up this progress for other projects. For example, it may allow vision systems to recognize objects and scenes without any supervision. Carl Vondrick, the first author of the finding, stated, “If you can predict the future, you must have understood something about the present”.

It is this mantra which companies hope to take advantage of for various other purposes, including financial investments. Researchers name deep learning as one of the most important, game-changing areas of advanced technology.

Deep Learning Algorithm

What is Deep Learning?

There is a lot of talk around this buzzword, but what exactly is deep learning? Deep learning is essentially the study of artificial neural networks and related machine learning algorithms that contain more than one hidden layer. This can be broken down in to its individual components. A single neuron might take in various inputs with assigned weights and output an answer. This could be connected with various other neurons to form a neural net. This neural net might have a complex web of differently weighted inputs and outputs it uses internally before generating an answer.

The difficulty in this task comes from deciding what weights will be used for each input. This is also called feature extraction when discussing what components of an image should be informative in making a decision. Normally, the onus of decision is on the programmer. The programmer is required to be insightful in order to code in proper heuristics for the computer.

Deep learning is a method to circumvent the challenges of feature extraction. This is because deep learning models are capable of learning to focus on the right features by themselves. This is done by training the models. Training involves showing the neural net a vast quantity of training examples. Then, with this data, the model iteratively modifies the weights to minimize the errors made on the training examples.

Deep Learning Algorithm

This may be more easily understood by means of an example. Imagine, that researchers want to teach a machine to recognize a set of handwritten numbers. For a standardized print set, this is fairly easy. The programmer simply needs to teach the machine one manner of representation. Say, Times New Roman font at 12 pt. However, things become trickier when using a set of handwritten numbers. There is such variability in handwriting that suddenly the machine cannot recognize which numbers it should categorize as a 1, 2, 3, etc.

For example, take the numbers 0 and 6. If the writer has certain quirks in their lettering and does not fully close the loop of the 0, the machine might mistake this as a 6. Hence, there is a need for the programmer of the machine to make decisions about the parameters and heuristics which outline when the machine should catalog the number as a 0 or a 6. However, it quickly becomes evident that describing the infinite possibilities which the machine might face is extremely difficult.

Deep learning overcomes this by having the machine teach itself the guidelines. The machine is given many training sets by which it teaches itself to distinguish the different numbers. With each success or failure, the machine adjusts the weights of each input and builds heuristics accordingly. Eventually, this means that the machine will be able to accurately recognize the difference between a 0 and a 6, despite all the possible variations.

Deep Learning Applications Today

Among the two leading groups in advanced deep learning research are Google and Elon Musk’s team. Google’s DeepMind and Musk’s OpenAI are at the forefront of deep learning applications. In 2011, Stanford’s Andrew Ng, a deep learning expert, joined Google X. A year later, the Google X team had reduced Android’s voice recognition error rate by 25% using deep learning. Google quickly followed up and began poaching as many deep learning experts as possible. This included acquiring Geoff Hinton in 2013 and picking up DeepMind in 2014 for $600 million.

One of DeepMind’s most famous achievements is creating the program AlphaGo. This was the first computer program to beat a professional human player of the board game Go. This is interesting because normal machines can beat humans at certain games such as Chess by brute force. Each move in chess has 80+ possible next moves. However, each move in Go has 250+ moves. It becomes exponentially more difficult for a machine to brute-force crack Go. Hence, AlphaGo was a demonstration of how training on 30 million moves of historical tournament data taught the machine to actually recognize complex patterns. This is very much a human trait, except on a higher level of operation.

One of the biggest problems, however, with machine-learning is that once a machine learns a certain task, it becomes useless at learning other tasks unless it overwrites the skills required for the one it knows. For example, a machine could not know how to simultaneously play Go and Chess. It would specialize in one of them. DeepMind succeeded in first breaking this boundary by using a technique called Deep-Q Learning. This reinforced the neural connections most necessary for the first skill while building connections for the second one on top of it.

By doing this, DeepMind succeeded in what has since become known as the Atari Challenge. This involves training the machine to play 10 different Atari games at once and measuring its proficiency at being able to play all of them. However, OpenAI challenged DeepMind head-to-head. OpenAI used a technique called neuroevolution to reach the same level of mastery at the Atari challenge in an hour. It took DeepMind a whole day. Furthermore, in the infamous walking problem, it took OpenAI 10 minutes compared to the 10 hours it took for DeepMind.

On a less theoretical level, Google is currently implementing DeepMind in various healthcare projects. It is collaborating with London’s Moorfields Eye Hospital. Moorfields has given the algorithm access to one million images from historical eye scans. The machine is training itself to read the scans and spot early signs which indicate degenerative eye conditions.

DeepMind is currently also running the Streams project. It plans to implement this in the NHS this year. Streams analyzes patient data and delivers cellphone alerts directly to a doctor or nurse if urgent intervention is required. Researchers trained it to detect signs of acute kidney injury. However, they could potentially teach it to spot many other conditions.

Deep Learning Algorithm

Deep Learning & Financial Markets

It quickly becomes obvious that the applications of deep learning are many and very exciting. One of the most interesting areas of deep learning application is that of finance. 40% of the world population is now online, and people use more than 2 billion smartphones every day. This is creating endless raw data for AI to process. This data is thorough in its encapsulation of our behaviors, interests, knowledge, etc.

All of this data is interesting to deep learning firms who hope to use this to factor in to making advances in financial technology. There were several false hopes and busts in the 90’s which have investors cautious of a miracle investment algorithm. However, the market is slowly starting to take in and utilize deep learning to improve financial returns.

There are a few examples of major investment firms which are starting to use deep learning to take on investment strategies. Bridgewater Associates has $150 billion in assets under management and recently started a new artificial intelligence unit led by David Ferruci. Ferruci led the development of IBM’s Watson. Renaissance Technologies, which uses deep learning-based technology claims to have returned +35% annualized over 20 years. Two Sigma Investments, which is notorious for using machine-learning algorithms for investment, currently has $32 billion under management.

Kevin Benedict, senior analyst at the Center for the Future of Work, writes: “We surveyed 2,000 executives across 18 countries for our Work Ahead report series and they predict AI will be the digital technology having the largest impact on their work by 2020… 46% believe AI will be critical to them within the next 40 months”.

Giant corporations are recognizing this movement and are aggressively poaching top AI teams. Twitter bought Mad Bits, Whetlab and Magic Pony. Apple bagged Turi and Tuplejump. Salesforce acquired MetaMind and Prediction I/O. Microsoft research chief Peter Lee compares the cost of acquiring a top AI researcher to the cost of acquiring an NFL quarterback.

Although these massive corporations are beginning to incorporate AI and machine-learning, one of the most interesting areas of deep learning is at the frontier of its technology. This can often be found with the plethora of startups that are currently focusing on using AI-based technology. Q1-Q3 2015 saw $47.2 billion invested in the global VC market. Of these companies, roughly 900 were working in the AI field.

In the UK, seed rounds related to AI and deep learning raised approximately $2 billion. 80% of the deals were below $5 million in size and 90% of the cash was invested into U.S. companies versus 13% in Europe. Some of these interesting startups include Narrative Science, which is using AI to read stock market data and compile reports which read as if written by humans. Another example is Kensho which compiles mass amounts of data and answers analytical questions.

However, it is evident that although this is a growing sector, financing and exit markets for AI and deep learning companies are still nascent. It is an area to keep a close eye on as technology develops and researchers find more groundbreaking ways to use deep learning.

I Know First’s Role in the Machine Learning World

I Know First is one of the Fintech companies which is using this technology to focus on the financial market. The company trains its machines on years of data to increase market profits for its clients. It is cited as one of Bank Innovation’s “5 Israeli Startups You Should Be Watching.” The company’s CTO Lipa Roitman developed this predictive system based on genetic algorithms. Open AI based its improved Atari Challenge machine on genetic and biological algorithms as well. It beat the record set by Google’s Deep Mind by 23 hours. Based on 15 years of historical data and current market data, I Know First’s system can identify patterns and predict future shifts in stock share prices over 6 different time horizons.

The basis of this is unsupervised learning. Each day, as new data is recorded, the system adjusts its own heuristics and learns from its successes and failures. Hence, I Know First’s system is constantly improving and becoming more accurate.

Conclusion

The market for deep learning and AI is definitely an exciting one. It is becoming increasingly important, and I Know First is strongly involved with pioneering the frontier of this development. Although it is still a nascent market, it is one that is set to rapidly grow in the next few years.