APPLYING MACHINE LEARNING IN THE U.S. POLITICAL LANDSCAPE: FORECASTING, POLLING, AND THE DOMESTIC SUPPLY CHAIN

Election Forecasting

The term “forecasting” refers to the ability to determine outcomes before they happen, and in this case, it is particularly applied to an ongoing election (Abramowitz, 2004). To forecast any election results most accurately, statistical modeling and analysis is likely required. This sort of modeling requires considerations for demographic data, previous election history, public opinion polling, and more (Abramowitz, 2004). This breadth of information can lead data sets to become very large and require extensive statistical labor. The next step for any decision maker would be to evaluate the results of a forecasting model to provide informed and actionable decisions.

1. Objectives: As previously mentioned, the industry of election forecasting has become more and more suitable to apply AI or its capabilities in some fashion. This can be attributed to the increased focus on statistical modeling. By utilizing machine learning, statistical labor in election forecasting can be reduced while providing increasingly informed and more accurate forecasts based on larger sets of aggregated data. Thus, the objective for this first analysis is to decrease labor required and increase accuracy of forecasts. Applying machine learning methods to the advanced modeling tools used to forecast any election provides a clear path toward improving the percentage of predicted outcomes. Between 2004 and 2016, 62% of expert election forecasts had accurately forecasted results for the presidential elections (Graefe, 2018). The potential of machine learning’s application can likely be a powerful force in driving this percentage of experts’ accuracy higher even if it is marginal. Marginal gains will become significant increases as any algorithm continues to discover valuable patterns to turn into actionable data. This improved data will allow campaigns to better apply resources and improve electoral win efficiency.

2. Technology/strategy: In any forecasting attempt, one must establish a time period in which the forecast is to be made and the amount of information that is available within that time period (Kaggle, 2022). Once these parameters are defined, one can begin to develop a forecasting model in traditional programs, such as Python (Kaggle, 2022). Enabling more advanced machine learning strategies throughout the development of algorithms for a forecasting model can potentially allow for larger sets of data to be processed more accurately. This has previously been applied in the private sector to forecast demand, and in many cases, it has allowed decision makers to avoid both tangible and intangible hurdles that may be faced. More importantly, in the use of machine learning in forecasting demand, managers have been able to reduce incorrect decision making due to former incorrect forecasts (Falak et al., 2022). Applying a particular program competitively for this analysis can be challenging as most robust programs for forecasting demand using machine learning techniques are entirely proprietary. In this circumstance, the application of machine learning algorithms in their most traditional sense to political forecasting models can likely reduce labor and provide more accurate forecasts.

3. Performance: In the case of forecasting elections, the clear key performance metric would be the percentage degree of accuracy in forecasting the correct election result. This information can be highly valuable to campaign managers, public officials, and others involved in any political unit. Nothing so far contradicts the potential that machine learning can be applied in the same fashion as it had been to demand in the private sector, using larger data sets to expand their parameters of forecasting while detecting and eliminating inaccurate forecasts. Effectively, this would enrich the data quality and the ability to produce a more informed and accurate forecast for any election decision, giving campaigns or political units a better perspective in allocating resources.

4. People: Regarding labor, an evident stakeholder would be the decision makers that benefit from advanced intelligence. The stakeholders also include programmers and data professionals developing and maintaining machine learning algorithms and forecasting models. Any decision maker or manager can utilize the information or decisions provided by a forecasting model. Based on its application in the private sector, there is evidence that the inherent quality of machine learning algorithms to use previous data inaccuracies to inform its future decisions would eliminate some degree of involvement by a programmer, processor, or manager.

Public Opinion Polling

Public opinion polling can be classified as a lot of things in the realm of the U.S. political system. But, particularly, this research sets specific parameters of how public opinion polling will be defined. For this analysis, public opinion polling will be classified as an independent, nongovernmental survey or questionnaire provided to a sample group with the intention of receiving information regarding the participants’ opinions concerning issues or elections (Pew Research, 2022). This information then can be utilized by political units, such as campaigns or interest groups, in order to make informed decisions about how to build policy, how to form an election platform, what demographics to target, and much more (Pew Research, 2022).

1. Objectives: Over the past few decades, the format in which polling has been conducted has remained relatively static as typical questionnaire and surveying tactics generally provide the intended results from a sample. Although the mode or medium through which public opinion polling is conducted has continued to evolve and has become increasingly reliant on web services and online data collection (Pew Research, 2022). This creates a gap for machine learning techniques and AI to be applied and enable this essential political function to be prime for growth in the technological sense. The evident objectives for this analysis would be to utilize aspects of machine learning tools to not only improve the accessibility and effectiveness at which the data is collected, but also enrich the ability to process that data and turn it into more refined results.

2. Technology/strategy: For this analysis, two particular machine learning programs were found to be especially applicable to the improvement of public opinion polling’s processing and medium on both ends of the spectrum. Initially, natural language processing (NLP) tools can be applied to the medium to improve data collection, effectiveness, and overall accessibility. Once data has begun to be collected, then machine learning principles can be applied through deep learning programs to process more unstructured data. NLP is a machine learning tool that is built to help computers communicate and interface with humans in their own language, allowing for spoken language feedback (SAS Analytics, 2022). NLP can be utilized for public opinion polling by allowing participants to submit survey answers using spoken language, increasing the overall accessibility of the entire population’s ability to participate. Promoting accessibility through NLP creates an even playing field for all Americans to have their say in any polling effort. Deep learning refers to a complex structure of algorithms and neural networks, specifically modeled after the functions of the human brain (Wolfewicz, 2022). Most importantly, deep learning allows for the processing of unstructured data, which, in this case, is especially valuable if polling were to utilize NLP for data collection. Deep learning programs can process images, sounds, or text in a brain-like fashion, utilizing its neural networks (Wolfewicz, 2022). In combination, these tools can enable decision makers to build better informed policy based on less structured and naturally spoken data. Both NLP and deep learning programs function separately and typically do not interface (Upgrad, 2020), but this approach would allow for pollsters to apply both technologies to improving polling in a variety of ways.

3. Performance: In this approach, as data is collected through means of NLP, it must then be processed and refined efficiently. Deep learning programs could then be utilized to draw conclusions based on unstructured, spoken audio recordings (Wolfewicz, 2022). Conclusions using deep learning are drawn using similar logical structure to that which humans would use (Wolfewicz, 2022), leading this analysis to believe its capabilities are applicable to processing audio data collection done through NLP-driven public opinion polling. The utilization of deep learning in polling projects has a much more significant impact on the performance metrics regarding the efficiency and accuracy of processing unstructured data sets as opposed to NLP having a more direct impact on the stakeholders involved.

4. People: Regarding the implementation of NLP to the medium of data collection, the impact is much greater on the metrics of the people to whom it applies. The primary stakeholders would be the U.S. citizens who are participating in public opinion polling. This group is in an especially good position to benefit from the use of NLP in public opinion polling as their accessibility increases exponentially through the ability to use spoken language. Effectively, this would allow Americans regardless of age, language, or disability to participate equally. Particularly, this can be helpful in accommodating circumstances through the use of a spoken question-and-answer format. This can already be achieved with audio recording, but the NLP machine learning programs are able to process this audio and convert it into meaningful data (Chapman et al., 2011). The utilization of deep learning will also have a degree of impact but on a different subset of stakeholders in the polling process: the pollsters and decision makers. Pollsters are the individuals who conduct or analyze polls (Merriam-Webster, n.d.), and the decision makers are those stakeholders previously described as having the ability to take action on the results of polls. Collecting more diverse and accurate information through NLP polling and being able to process it through deep learning applications will certainly provide a level of increased insight for pollsters and decision makers.

Vote by Mail

Vote by mail refers to the processes by which U.S. citizens are allowed to cast their vote in a state or federal election through a mail-in ballot. This ballot is delivered to the voters’ home, filled out, and then shipped back to the corresponding Secretary of State (USAGOV Contact Center, 2022). Many states extend a vote-by-mail option to all its eligible population, whereas other states require a request, and these are considered absentee ballots.

For this context, any reference to vote by mail will include absentee ballots and all mail-in ballots that would be processed for a state or federal election (League of Women Voters, 2020). Vote-by-mail procedures inherently require cooperation from the USPS and, in turn, directly participate in the greater domestic supply chain.

1. Objectives: Vote-by-mail services rely inherently on the capabilities of the U.S. domestic supply chain. Ensuring the optimization and overall functionality of supply chain procedures could serve as an evident path toward improving vote by mail on a state and federal level. The objective would be to source and implement a machine learning tool that can directly improve the efficiency of vote by mail processing in both regional and national USPS facilities. Information regarding the techniques utilized to process mail-in ballots is strictly proprietary; thus, this analysis focuses on projections regarding the application of known machine learning tools.

2. Technology/strategy: The term “processing” regarding vote by mail can be defined differently for different states. One commonality that reaches across the union is the requirement to compare the signature on the outside of the ballot with the corresponding voter’s signature. This is done to ensure a match and the legitimacy of the ballot (National Conference of State Legislatures, 2022). Vote-by-mail procedures could implement visual inspection through an AI-driven camera stream analysis. This analytical tool has been utilized to spot errors or inconsistencies in large production facilities (Helo & Hao. 2022). It can be trained easily using images and can effectively match signatures nearly automatically (Helo & Hao, 2022). Utilizing its neural networks, an AI-based camera stream analysis could match visual data and establish the necessary linkages to massive sets of statewide or nationwide data, such as a voter signature repository.

3. Performance: Implementing AI-based camera stream technology would likely substantially improve the speed at which mail-in ballots can be processed as this technology would enable the initial processing (signature matching) to be completed while the singulated mail is still on the conveyor belt. Singulation refers to the process to create a single-file nonstacked flow of ballots so that a top read and bottom read camera system can adequately see the signature (Costanzo, Mark B., personal communication, November 2022). This eliminates the need to physically process and organize the mail-in ballots to then be authenticated by an employee or scanner. Eliminating extra unnecessary steps is a simple way of improving the overall efficiency of the supply chain. In circumstances such as presidential elections in which mail-in ballots are increasingly more popular, eliminating processing steps could lead to substantial returns in time saved and potentially produce election results faster. The specific technology can also improve the overall accuracy of the signature match and most certainly reduce the potential for human error or intervention by bad faith actors.

4. People: The processing of mail-in ballots requires a couple of steps; the first of which has been established as the signature verification. This is typically handled by an employee who must manually verify the signatures or run them through a scanner (National Conference of State Legislatures, 2022). These employees would be the primary stakeholders. Ballots must then be opened, prepared for tabulation, and stacked accordingly by the same employee (National Conference of State Legislatures, 2022). This turns the processing of mail-in ballots into tens of thousands of hours of labor that requires a large employee base per each state. In the case of national or presidential elections, these labor hours can increase dramatically. Implementing the AI camera technology into the supply chain’s operational line can remove the initial processing step and potentially reduce the necessary labor hours substantially.