Predicting Economics with AI

Last updated Mar 20, 2020 Edit Source

• Seo
• Blog post

Certain assumptions about economic models just don’t play out as you expect in the real world. These assumptions include: perfect competition, infinite buyer knowledge, and endless resources for production, and they form the basis of our modern economic system.

The stock market, for example, does not reflect real economic activity, as much as it does consumer sentiment. But with the advent of Big Data technologies and machine learning, can we possibly overcome the barrier of unrealistic models and lack of foresight into the future economy by leveraging the vast amounts of data available about consumers?

But what if the Federal Bank got smarter? They currently have very little actual visibility on how the market is performing and whether we are heading for a recession. It is strange to some people that within some companies you can ingest data from several different sources and automatically forecast the size of your revenue pipeline to the thousand dollars, yet very few of the thousands of highly trained hedge fund managers can reliably beat the market. Perhaps this suggests that luck and randomness are under-appreciated factors in the economy.

To the surprise of some of us, AI is alive and well in our economy today and is actually driving a large sector of it. As of this article’s publishing in December of 2017, 60% of the daily trading volume of the NYSE is executed by computers in what is called high-frequency trading. In short, high-frequency trading leverages the fastest internet speeds and computer algorithms to capitalize on small price differentials that only exist in the market for fractions of a second.

Indeed, algorithms and automated systems make up a large portion of the agents trading on a given day outside of high-frequency trading. For example, the subreddit Algorithmic Trading is a vibrant online community where people are creating computer algorithms to trade in stock exchanges. Unsurprisingly, these algorithms are more effective than human traders because they do not sleep, have no emotion, and can make instant decisions, but they can be costly to build because of the expertise needed.

In the business world, Quants and other financial firms have been using computer technologies and statistical methods to execute transactions on large scales for decades. Blair Hull of Hull Trading Group sold his company to Goldman Sachs in 1999 for pioneering the use of computer-assisted trading. He also hosts a yearly competition, which is in its second year as of this article’s writing to predict the returns of the S&P 500 using several market and associated metrics. The data used has interesting features for those looking to get a flavor of what is used in predicting market returns, but the competition is currently only available to UCSB students.

There are several difficulties in building compelling systems to predict the stock market:

1. The market consists of millions of agents, each with their own interests who individually would be difficult to predict.
2. We can show mathematically that stock prices cannot be inferred in the future from past values. One can show that because the value of an asset in the stock exchange can be modeled as a Brownian Motion, the mean of the expected value of the market is the current value, so we can never expect it to go up or down based solely on past information.
3. Even if we do build an effective system to beat the market, it is likely that over some undetermined period, the conditions of the market will change and the algorithm will no longer be effective. For example, an algorithm that thrived in the bull market last month may likely fail to adapt to this down turning market in the wake of the COVID-19.
4. Most importantly, the market reflects consumer sentiment the most out of all other factors. Consumer sentiment here means the overall distribution of how active members of the economy perceive a particular topic – usually positively, neutrally, or negatively – such as the oil industry.

Now that we have insight into consumer behavior that we have never had before, we are, in theory, capable of using AI to refine behavioral economics and actually predict economic scenarios before they occur[^1]. While this sounds great, the size of a comprehensive project such as this may be staggering. While an AI project to make our economy more efficient and predictable would have huge monetary implications, the scale and complexity of such a system make it unlikely to be built anytime soon. [1^] https://towardsdatascience.com/how-ai-will-redefine-economics-ec305e3cb687?gi=f1e16cae4cf2

Recently, the CEO of Starsky Robots – an autonomous vehicle company who aimed to build Semi-trucks – announced that they will be closing up shop indefinitely. Their initial mission statement was strong because the potential upside for automating semi-trucks is huge. But Starsky Robots was forced to close after 3 short years working towards this goal, which raises the question of what happened. Was the goal post moved, or was it too far away, or just not worth the effort?

Stefan, the ex-CEO of Starsky Robots, argues that the most difficult part about building a profitable autonomous vehicle is safety. However, this was not exciting enough for investors for how much time it took to build. While Tesla is getting press coverage for automatically changing lanes, Stefan explains that it will take a dedicated team of engineers 10 years to build a sufficient autonomous vehicle system because of the complexity of the task and consumer safety concerns and that the money is not there for this kind of team.

Similarly, the concept of AI that finally gives us oracle powers over the economy is a large reach. We have already discussed the impact of algorithmic trading, but there are limits to what it can accomplish. With such a complex system as a global economy, you could have 1,000 teams of the smartest people build models to comprehend the economy and get 1,000 different systems that lack consensus. While the economic impact would be great if AI existed that could accurately predict recessions, or take the role of Jeremy Powell; this AI would have to be so large and so complex. How could we trust such a complex AI? And is the money even there to build a trustable Economic AI overlord? In fact, transparency is a crucial factor in evaluating AI. (Check out our blog post on Model Transparency and Explainability here)

Top teams of the smartest finance wizards are already building systems taking in information about trades in the Baltic Sea, consumer sentiment from Google, and sentiment from economic articles to perform in stock markets. Algorithmic trading is a tough scientific and engineering problem, and a lot of smart people are trying to make money by doing it correctly. While the potential reward in building AI that understands the economy is great – as shown by the story about self-driving semi-trucks from Starsky Robots – there is not much sexy about safety, and if we don’t understand or believe the AI we built to predict the economy, then who would use it? There are just too many uncertainties to be solved.

If you would like to understand how to leverage the power of AI to solve real problems that affect your business and that you can trust, request a demo or contact us today to start a free trial.