Deep Learning in Finance
Siri is the voice controlled AI behind most Apple products. It can recognize your speech, analyze your sentiment, and answer questions. The man who was key to its development, Babak Hodjat, is now Chief Science Officer for a new fund called Sentiment.
Hodjat is currently experimenting with applying deep learning to trading. Just as deep learning can pinpoint particular features that show up in a photo of a cat, he explains, it could identify particular features of a stock that can make you some money.
While funds have famously used computers for high speed trading, this is something new. Instead of concentrating on the millisecond inefficiencies present in markets, deep learning is being used to create strategies over longer periods of times — hours, days, or even weeks.
How the system works is a bit like the sci-fi movie Tron. To start, it randomly generates a series of artificial traders and compares their performance over historic data. The best ones from each series are used as the basis for a new set. It continues the process, discovering the artificial traders capable of performing on their own. These are the ones it keeps. All the others are destroyed.
Before moving forward, we need to take a step back into how most funds use computers for trading. Computer programs are made up of algorithms. An Algorithm is nothing more than a set of instructions. They are core to writing programs — similar to a recipe. While human cooks will eventually memorize how to make the dish, computers do not. They need the set of instructions each time.
A trading algorithm is a set of instructions for a computer to use in determining when and how to make a trade. For example, a simplified version tells the computer if ABC goes down in value by $3 then short it. If ABC goes up in value by $2 then take a long position. There are other instructions for when to close the position and limits to the amount risked. However, all a trading system does is follow instructions.
In contrast, a deep learning system is based on learning data and not instructions. For example, Google’s DeepMind learned how to play and beat Atari games with only a few pixels as input. All with no instruction.
Originally, deep learning was to be modeled on the biological brain. However, researchers found that by augmenting natural neuron models with things like backpropagation — passing data in the reverse direction, they could accomplish improved learning abilities with fewer nodes.
Today, deep learning is made from neural networks with multiple hidden layers between the input and output. The issue is that most, if not all, of these systems are domain specific. Making it similar to other applications. So far, the goal of a universal machine has not been accomplished.
Facebook’s deep learning can determine who is in each picture uploaded to their network. Google’s system can beat professional players at Go. IBM Watson won at Jeopardy, a US Game Show. However, none have yet to pass the Touring Test. A measure of an AI system’s ability to mimic human speech, emotions, and language patterns by chatting with a person. To pass, the system must be indistinguishable from a real man or woman.
While very difficult, it will probably happen within the next few years. For now, deep learning is limited to specific areas with clearly defined situations such as Google Duplex making appointments or IBM Assistant directing customer service requests.
The truth is that AI research began in the 1950’s with laboratory systems learning chess strategy. At the time, those involved with the study believed we would have complete artificial intelligence by the 1970’s. What occurred was they failed to appreciate the difficulty present in the next steps. These are the challenges we are trying to solve today.
Value or Technical
In human powered stock trading, the core areas are value investing and technical analysis. Value Investing is credited to British Economist Benjamin Graham. Later popularized by Warren Buffet. This camp‘s focus is on the asset more so than the market. Technical Investing is using charts of streaming market data to make decisions based on the market and nit the asset.
Value Investors look for stocks that are trading below the value of the company they represent. For example, the stock for a mens clothing company is trading at $10 per share with 1 million shares in circulation. A Value Investor considers the company is selling for $10 million. The next step is to determine how much that company is actually worth right now. Answering the question, “Would I buy this company for $10 million?” If yes, then they purchase shares. Otherwise they pass.
Once a Value Investor makes an investment, they hold it until given a valid reason to sell. Their outlook is long term as their goal is to buy companies. This works the same for one share or thousands.
In contrast, Technical Analysis only considers the assets market price. The only question they attempt to answer is will this asset go up or down in the next n timeframe. They do this by looking at charts.
The Technical Analysis crowd looks to close their position as quickly as possible since holding the asset represents risk. Strangely enough this is the most popular form of self-guided investing. Probably due to the fact that brokers make far more from active traders.
No one can argue with Warren Buffets success with Value Investing. However there are funds and individuals who make large sums from technical analysis too. The problem is that it’s most likely impossible to determine if an exchange traded asset will move up or down on a consistent basis. Yet it seems to be possible.
Fooled By Randomness
A fair coin will trend toward a 50% probability of heads or tails each time it is flipped. Sometimes, it will have a string of heads. Others, it will show tails for several turns. However, the more it is flipped the more likely the 50% distribution becomes.
We all know this, yet people are often fooled by randomness. Gamblers Fallacy is the concept where people erroneously believe they can predict a random event. Consider the coin showing heads on the past three events. Most gamblers will bet that the coin will come up heads on the next attempt. Yet it still has the same 50% chance of being heads as it does tails.
Random events have no memory. It does not matter what the last result was. The current event is new and offers a known probability for success that is unrelated to the last. So how do Technical Analyst win at this game?
Just as occurs in turns of a roulette wheel, there are periods of winning. One can look at a chart and jump on a trend. Seeing a trend in progress is easy. Knowing when it will end is the hard part. That may be the moment you take a position or a few days later. However, it will end.
Sure there are theories on the wisdom of crowds and times when things are more predictable, but overall it seems that markets are almost efficient and prices are almost random.
Artificial Intelligence, no matter how advanced, will never be able to predict a truly random event. Perhaps it will discover how to take advantage of the almost random to improve predictions similar to how human traders do during a Federal Reserve announcement.
For example, during the start of an announcement, trading for the affected asset, such as oil futures, will move in one direction very rapidly. At this moment, direction is not random. The random part is knowing when a majority will exit their trade and reverse the market movement.
There may be promise in using Deep Learning to actually discover how random exchange trading actually is. If it could find patterns to indicate the moments when pricing is predictable then it would forever change investing. That is if these moments even exist.
The seemingly best use of deep learning in trading will most likely come in the form of Value Investing. Where networks train to spot companies trading below their market value and creating estimates to when the market will catch on. That way, the system can make trades using the highest probability of success.