The total size of all global data hit 20 zettabytes in 2017. For 99% of people, that number probably means nothing, so picture this: if every 64-gigabyte iPhone were a brick, we could build 80 Great Walls of China with the iPhones needed to store all the world’s data.
We are awash in an ocean of data that grows bigger by the second. And it’s a complete and utter mess.
90% of web data is unstructured, meaning it’s in a format that cannot be easily searched and understood by machines. Poor data quality costs the US economy $3.1 trillion a year according to IBM (IBM). We have become a society that is excellent at producing, storing and sharing data, but we’re lousy at making it useful.
Poor data quality represents the single largest hurdle for developing useful artificial intelligence. It doesn’t matter how “smart” machines become if they’re fed data that is inaccurate or incomprehensible.
The Size of the Data Management Problem
Poor data quality is a familiar problem for those who analyze data for a living. A recent survey found that 60% of data scientists devote the majority of their time to cleaning and organizing data, as shown in Figure 1.
Figure 1: Cleaning Data Takes the Most Time
Comparatively, just 9% of data scientists devote the bulk of their time to mining data for patterns. Cleaning and organizing data has become such a big task that it leaves precious little time for analysis.
When people predict that AI will make human workers obsolete anytime in the near future, they are ignoring the data quality problem. AI and machine learning may be able to replace those 9% of data scientists who are mining data for patterns, but it will still need the 80% working on collecting, cleaning and organizing data.
More importantly, data scientists need to re-orient their thinking around data quality. More time needs to be spent upfront and on collecting data in a high-integrity manner rather than retroactively “cleaning” it. We are not sure that it is possible to retroactively clean data enough to meet the needs of useful AI. If you cannot validate the data back to its source, then how do you know it is clean? And, if you are going back to the source to validate, then you might as well collect it from the source.
Gradually, leaders in AI seem to be understanding that their job is as much, if not more, about data management as it is about new technologies. Facebook (FB) just hired Jérôme Pesenti, the former head of IBM’s Big Data group, to run its AI efforts. Pesenti replaces Yann LeCun, who will now focus on his core expertise of research.
While research into deep learning and neural networks makes the most headlines, Facebook understands that data management is crucial to delivering bottom-line business improvements.
AI Has a Long Way to Go
Once you understand the limitations imposed by poor data quality, the challenges facing AI become much clearer. Successes like the Go-playing computer that beat the world champion are misleading because they are enabled by completely structured and easily interpreted data points. Accordingly, the subset of tasks that AI can perform effectively remains small, according to LeCun:
“In particular areas machines have superhuman performance, but in terms of general intelligence we’re not even close to a rat.”
Other researchers have faced similar hurdles. Machines can handle finite and discrete data points well, but even a minor degree of ambiguity can trip them up.
“We are still a long way from computers being able to read and comprehend general text in the same way that humans can.”
The last quote comes from Microsoft (MSFT) CTO, Kevin Scott, in a LinkedIn post celebrating the development of an AI that could read and answer questions about Wikipedia pages at the level of an average human. Despite its success, the machine struggled when asked to go beyond simple facts and make intuitive, but still logical, leaps. Contextual clues that a human would understand easily are still incomprehensible to machines.
This problem is even more pronounced when it comes to using machines to read financial filings. One of the biggest issues people face when trying to use natural language processing on financial filings is that the language in these documents is far from natural. If machines can get tripped up by Wikipedia, imagine how they respond to the jargon and legalese that fill your average 10-K.
In the financial world, as in the rest of the economy, AI has had its biggest successes working with data that is structured and standardized. JPMorgan Chase (JPM) has automated hundreds of thousands of hours of work annually by developing machine learning tools to read commercial-loan agreements. These contracts are standardized, which means the machine only has to navigate minor differences.
While machines are replacing humans in these rote tasks, they struggle to make the leap to more sophisticated analysis. Rather than try to tackle the complex task of reading and analyzing financial filings, most applications of AI in the investing business focus on mining patterns out of trading data, alternative data sets, or sentiment and other non-financial indicators. As we discussed in our first piece in this series, these efforts have not yet led to superior returns. It’s been easier to apply existing AI and machine learning tools to new data sets than it has been to teach AI to analyze old data sets, especially data directly from SEC filings.
Structuring Financial Data: Not as Easy as Most Think
In theory, financial data in filings would be more structured and standardized, or we could make it that way easily. We have centralized bodies (FASB, SEC) that govern financial reporting standards. Public companies employ teams of accountants and lawyers to conform to these standards.
In reality, the data remains highly unstructured and variable, and we expect that it will only get worse. The most prominent effort to make financial data machine readable, XBRL, remains riddled with errors 10 years after its initial deployment. While companies are required to submit XBRL filings, they’re not required to verify them, and only 8% of companies carry out voluntary audits. Until XBRL is strictly enforced by the SEC, it does not stand a chance at being reliable.
Without SEC enforcement, most companies will continue to under invest in the resources necessary to get their XBRL filings right. They will either leave the task in the hands of accountants that lack the technical expertise to do the job, or they outsource the job to a third-party that doesn’t fully understand the company’s financials.
Again, we come back to the disconnect between people with technical expertise and those with subject matter expertise in finance.
As long as XBRL and other efforts to structure financial data are treated as curiosities that companies can safely ignore, it will be almost impossible to make meaningful progress on this front.
Structuring Data: More About Team Than Technology
As long as financial data remains unstructured, existing machine learning tools effectively cannot process it. Meanwhile, the cost of employing the highly-trained analysts needed to manually structure data remains prohibitive.
Our solution is to leverage our deep financial expertise into software that enables highly-trained analysts to structure data with unrivaled efficiency. In essence, we arm human subject matter experts (SMEs) with technology at every step of the data collection and modeling process. For example:
- Analyst and programmers work together to build the AI. Analysts and programmers anticipate and address problems from multiple perspectives from the outset. The ability to communicate clearly between financial and technical experts is critical to building machines that work. Precipitating potential problems at the start, along with frequent iteration and joint and rigorous testing, helps build machines that robustly do something small. One step at a time, we build on technology that we know works. However small that step may be, each step means less work for human SMEs. We don’t send programmers or data scientists to analyze the data in isolation.
- During data collection, human analysts leverage a multitude of sophisticated algorithms to validate the data points collected by the machine. We use our big data experience and financial expertise to automatically identify data that’s potentially wrong – from values that are too big or too small, to data points that show up in the wrong places, to data relationships that don’t make financial sense – and direct analysts to make decisions about how the data should be collected. We also track how analysts answer questions raised by the Robo-Analyst when it is not 100% sure.
- The machine also tags items it doesn’t recognize so that SMEs can help. If the machine is not sure or does not have adequate precedents for making a decision, analysts are called in. This process frees analysts from banal, mundane work so they can focus on new and more difficult problems.
- All the verification work and new data collection done by human analysts is fed back into the machine so that it is constantly learning how to better identify and interpret the data points it finds. To maximize the efficacy of our machine, we aim to give it as much information it needs to correctly analyze filings. To teach a machine well, we need to think like machines. Accordingly, every data point in the 120,000 filings parsed into the machine by human SMEs is associated by 10+ pieces of information. These pieces of information include data value, the associated text, the location in the filing, and many other features that are taken for granted by most humans but provide critical context for the machine.
- Our data collection process includes sophisticated corporate performance and valuation modeling of the data, against which we can automatically run multiple quality assurance algorithms to flag modeled results that are unusual or have been linked to errors in the past. We leverage our financial expertise to make sense of the data and flag instances where the results might not be right. This process adds significant integrity to the underlying data compared to data collection by humans who are not experts in accounting or finance or do not have a model to help them analyze the impact of the data they collect.
- The scale and efficiency of our process has a virtuous effect on our data validation processes. The more models we can build, the more potential data anomalies or errors we can find and feed back into the machine. The more we do, the more we can teach the machine, and, in turn, rely on it to do more. This approach gives us significant advantage over systems or analysts who can only view a few models at a time.
Working with machines presents many new challenges to our society. It is not something we’ve done before and, not surprisingly, we have a lot to learn, and so do the machines. One thing we know for sure is that people that are best at teaching machines will have the best machines, and the people with the best machines will control the future.
This article is the third in a five-part series on the role of AI in finance. The first, “Cutting Through the Smoke and Mirrors of AI on Wall Street” highlights the shortcomings of current AI in finance. The second, “Opening the Black Box: Why AI Needs to Be Transparent” focuses on how transparency is crucial to both developers and users of AI. The last two articles will show how AI can lead to significant benefits for both financial firms and their customers.
This article originally published on January 26, 2018.
Disclosure: David Trainer and Sam McBride receive no compensation to write about any specific stock, sector, style, or theme.
Article by Sam McBride, New Constructs