Public policy. State legislation. Court jurisdiction. Upon hearing these words, some associations that may pop up are perhaps the hours spent trying to digest the dense language of political documents for a history assignment, or just general frustration with the unpredictable nature of politics. Having spent much time finding and analyzing information on different pieces of legislation, Tim Hwang ’14 from the Woodrow Wilson School has much experience in dealing with both. Instead of wallowing in the difficulty involved in the process, however, Tim Hwang has taken matters into his own hands.
As CEO and founder of FiscalNote, a real-time government analysis startup, Tim Hwang has redefined the process of evaluating legislation. Currently, whenever people want to analyze and see the effects of a bill, they need a large staff to track and examine the legislation, a lot of time, or both. FiscalNote aims to change all of that. Analyzing state and local legislation, FiscalNote not only gives real-time summaries of the over 200,000 bills proposed across state legislatures as they go through the process of becoming a law, but also predicts whether the bills will pass. With such capabilities, a few questions arise: What secret political formula is FiscalNote using? What army of politically savvy geniuses does FiscalNote employ? How is FiscalNote able to do so much so quickly? The answers, however, do not lie in any magical political theory or group of politically gifted people. Instead, they lie in the realms of computer science and statistics—machine learning.
At its core, machine-learning algorithms are programs that tell a computer how to learn a certain task rather than directly telling it how to perform the task. For example, it would be nearly impossible to manually tell a computer how to play chess, for there are innumerable possibilities that could occur. However, if we give the computer information about lots and lots of chess games, the computer could eventually develop a strategy for playing chess by observing the moves made and the outcomes of those moves. While machine-learning techniques are by no means 100% correct in what they predict, they are effective enough to help handle large amounts of data that would be tedious to look at manually. In a similar way, FiscalNote “teaches” its software how to find and analyze information on legislation by using various machine-learning techniques in three stages: finding, classifying, and predicting.
In the “finding” stage, FiscalNote gathers information on state and local legislation by enlisting web crawlers, programs that scan the text of the desired websites and then record the relevant information. For example, a web crawler could look at all of the articles on the New Jersey state website and then save the text of any of the articles that contain the words “bill” or “law.” By building a web crawler that scans various public websites, such as state and local government websites, FiscalNote can constantly update information on new pieces of legislation.
With the information it finds, FiscalNote then classifies the legislation it finds with the different industries, such as healthcare or education, which the legislation may affect. To do this, FiscalNote uses two core subcategories of machine learning: natural language parsing and topic clustering. Natural language parsers are essentially programs that determine the “meaning” of a piece of text. Topic-clustering algorithms, on the other hand, divide a set of information into different categories. For instance, given a random article, the FiscalNote natural language parser could first determine the name, authors, and description of a bill while the topic-clustering algorithm could determine that the bill is associated with the health-care industry based on key words such as “doctor,” “medicine,” and “hospital.” Using a combination of the two, FiscalNote is thus able to determine what a piece of legislation is about and which industries the piece of legislation is likely to affect.
What sets FiscalNote apart, however, is its final stage—“predicting.” By looking at thousands of different factors, such as the legislators’ demographics, political views, and previous decisions, FiscalNote can actually predict with greater than 90% accuracy whether the bill in question will pass. At first, the original algorithm took approximately six weeks to finish. However, after much fine-tuning, the whole process now takes only a couple of minutes to go from start to finish, allowing FiscalNote to provide real-time updates on bills as they become law.
With its capabilities, FiscalNote can have far-reaching impacts on the relationship between business and government. With FiscalNote, businesses will no longer need a large team to analyze new legislation, and with more information on state and local legislation, businesses will have more confidence when developing their strategies. Furthermore, FiscalNote makes a government’s actions more transparent by giving businesses easily accessible and clear information on the government’s legislation, potentially changing the way governments interact with businesses. Outside of using FiscalNote in a business context, Tim Hwang is also considering offering FiscalNote’s services to universities for studying state and local governments, providing more comprehensive analysis of bills, and expanding FiscalNote’s reach into regulations, court cases, and speeches as well.
Although politics and machine learning seem at first to be on different ends of the academic spectrum, Tim Hwang has combined the two to create powerful results. Essentially, with machine learning, he has “taught” a computer to do research and analyze legislation, something that even many people have difficulty doing today. The power in machine learning lies in its ability not to do things correctly, but to do things correctly most of the time. Tim Hwang best summarizes the credo behind machine learning when he says, “Even if you don’t produce 100%, even if you’re 80% of the time right, it’s better than guessing.”