“Trading Floor”. What do you picture today when you hear these words? You might think of men in suits frantically gesturing and incessantly cursing at each other or a similarly chaotic environment. However, these once ubiquitous floor brokers are becoming replaced by high-speed computer programs.
For example, Citadel Securities trades 900 million shares a day (this accounts for 1 in every 8 stock trades in the US). Only 40 people work on the trading floor of the firm, overseeing computers that employ algorithms to fill stock orders. Goldman Sachs employs more programmers and engineers than Facebook. $40 billion was raised by financial technology (fintech) companies in 2018. If it wasn’t already clear, technology will disrupt the financial sector. Artificial intelligence is one of the technologies spearheading this change. Before we can understand AI’s applications to financial services, we must understand the technology itself.
Machine learning, a subset of artificial intelligence, focuses on developing computer programs that autonomously learn and improve from experience without being explicitly programmed. The three broad types of machine learning are supervised learning, unsupervised learning, and reinforcement learning.
The goal for supervised learning is to create predictive models. Initially, a training data set with labeled input and output examples is fed to the algorithm (hence the name supervised). Then, the algorithm runs on the training set with its parameters adjusted until it reaches a satisfactory level of accuracy. From this analysis, the algorithm creates a function that can predict future outputs. In the image above, the AI model is given pictures of cats that are labeled as “cats”. The model is then trained on the labeled data of cats until it can recognize the patterns in the images of cats. As a result, the model would be able to predict if later images are showing cats or not cats by responding to the previously recognized patterns.
The goal for unsupervised learning is to find patterns in data. Contrary to supervised learning, an unsupervised algorithm is given a training set without classified or labeled examples (hence the name unsupervised). To discern patterns, the algorithm uses clustering. Each cluster is defined by the criteria needed to meet its requirements; that criteria is then matched with the processed data to form the clusters. The training set is then broken into clusters based on common features. In the image above, the input data has no class labels and comprises of fish and birds. An unsupervised model built using this input data will create one cluster of fish and another cluster of birds by grouping the data based on common features.
The goal for reinforcement learning is to train a model to make a sequence of decisions that will maximize the total reward. In reinforcement learning, a machine learning model faces a game-like situation where it uses trial and error to solve the problem it is facing. The programmer manipulates the model to act in a certain way by adding rewards and penalties. As a result, the model is incentivized to perform behaviors that have rewards and discouraged from performing behaviors that incur penalties (this feedback is the “reinforcement”). Once the model is left on its own to figure out the best approach to maximizing reward, it progresses from random trials to sophisticated tactics. For example, Google’s Alpha Go computer program trained to play the game Go and ended up beating the world champion. This was a huge achievement because there are 10¹⁷⁰ possible board configurations (more than the number of atoms in the known universe) and no computer program had previously beat a professional Go player.
Natural language processing is another subset of artificial intelligence with uses in finance. The overarching goal of natural language processing is simple: decipher and understand human language. Speech recognition software (ex. Siri) isolates individual sounds from speech audio, analyzes these sounds, uses algorithms to find the best word fit, transcribes the sounds into text. After converting the natural language into a form a computer can understand, the computer employs algorithms to derive meaning and collect essential data from the text. Now that we understand machine learning and natural language processing, we can look at artificial intelligence in finance with a better understanding.
Artificial intelligence has several diverse applications on both the sell side (investment banking, stock brokers) and buy side (asset managers, hedge funds).
- Firms are using machine learning to test investment combinations (credit/trading)
- Banks are experimenting with natural language processing software that listens to conversations with clients and examines their trades to suggest additional sales or anticipate future requests (credit/sales)
- Banks are using machine learning algorithms that recommend the best rate swaps for a firm’s balance sheet (rates/trading)
- Client messages in inboxes and electronic platforms are monitored by natural language processing software to determine how they want to allocate large trades among funds (rates/sales)
- Supervised machine learning algorithms seek correlations among asset prices and other data to predict currency prices a few minutes or hours into the future (foreign exchange/trading)
- Reinforcement learning AI runs millions of simulations to determine the best prices to execute client orders with a low market impact (cash/trading)
- Natural language processing software can read contracts and notify clients of swap expirations and other terms (derivatives/sales)
- Computers are sifting through historical data to identify potential stock, bond, commodity, and currency trades, using machine learning to project how they would perform under various economic scenarios. Historical data is also examined to assist in setting the size, timing, and duration of wagers (identify trades/portfolio construction)
- Machine learning algorithms analyze data on market changes to accordingly model changes to trades. Furthermore, analysis is performed on valuations and prices are forecasted (monitor trades)
- Algorithms analyze diverse sets of data such as consumer sentiment towards brands and oil-drilling concessions. Data such as satellite imagery and property listings can be used to track economic trends. Natural language processing also analyzes transcripts of earning calls, reads the news, and monitors social media. Commentary from central banks and conferences are also analyzed for keywords and sentiment (ongoing research)