User behavior prediction using AI. A definitive challenge

This time, we’re using AI and Quantum UX as an experiment for user behavior prediction. Please note that this is just an experiment, and we don’t have many resources (or time), but we believe it’s an intriguing experiment that, if successful, can reveal the true potential of AI as a predictive tool, and more importantly, the sheer power of Quantum UX.

User Behavior Prediction: Introduction

User behavior prediction always occurs within a context. This context is a blend of existing conditions and the environment, along with how these conditions influence user behavior in that environment (or vice versa, how the environment shapes behavior). To provide a simple example: if a user desires a hotdog in a fancy restaurant, there are at least two possible scenarios:

1. The restaurant doesn’t serve hotdogs, and the user must choose something else. In this case, the conditions shape the user’s behavior.
2. The user realizes that fancy restaurants usually don’t serve hotdogs and adjusts their behavior accordingly. Here, the environment provides clues, and the user alters their original behavior to adapt.

With this in mind, let’s discuss our environment.

As many of you know, our main offices are located in Argentina. At the time of writing this, we’re approaching the primary elections, leading to many predictions concerning the results.

One thing that caught our attention was the elections in previous weeks and months in some provinces. What was most striking was that opinion consultants’ forecasts in those elections were far from reality.

This prompted us to think: why not conduct a small-scale experiment? This will prove whether our approach is valid. If not, it will highlight areas for adjustment and what we might be overlooking. While we are based in Argentina, we do not work for Argentina, ensuring our complete neutrality. Although each of us has personal political views, we are not paid by any candidate or media, so our only interest in the outcome is the validation of our methodology.

The Methodology to Measure Public Opinion Using AI

One thing we noticed in the aforementioned studies was a methodology limited in both the number of cases and geographical dispersion. We lack sociocultural or economic information about the demographic groups studied, but most were conducted in Buenos Aires, our country’s wealthiest city. This fact alone means nothing, since there are impoverished areas in the city, but the geographical selection is important nonetheless.

This realization resonated with us, and we saw an opportunity to attempt a different methodology. As a result, we initiated a modest experiment using Artificial Intelligence, Statistics, and Quantum UX.

We began by defining the group to study and the methodology to acquire insights. We wanted as many people as possible, as widely dispersed as possible, and with diverse demographics.

Reaching the number of people we desired through conventional methods (phone, street surveys, in-house visits) was unrealistic due to our limited resources.

We opted for a simple plan: entering generic or thematic groups related to our country on various social media platforms, including Facebook, Twitter, WhatsApp, Twitch, and Instagram. A crucial element was avoiding any group or personality with pre-existing biases.

Once we defined this methodology, we assigned five researchers to follow a single social media website for a full week, from July 31, 2023, to August 6, 2023. Elections are scheduled for Sunday, August 13th, and an electoral ban starts today, so no new information on political parties’ voting intentions can be published. This is why we also release a somewhat incomplete report.

The methodology itself was uncomplicated: each researcher joined discussion groups or news feeds and simply added a number “1” to each candidate mentioned by a commenter. Main posts were ignored; we only considered comments (e.g., reactions to the post).

Additionally, we only considered candidates that we thought might advance to the main elections in October. This added a small percentage of uncertainty, which was factored into the final results.

One thing to note: we are not inexperienced in this matter. We previously helped a new party become the fourth political force some years ago, and we managed to get them two seats in the national legislature and several minor positions in various cities and provinces. Interestingly, this party has lost most of its relevance and was not considered for this experiment. Nevertheless, it’s important to clarify this since part of the analysis was qualitative and conducted by professionals with experience in the field, which obviously influenced the results.

Lastly, to explain the voting system, we must say that it is based on the selection of candidates who may or may not be part of the same list. The candidate who obtains the most votes within the same list wins the internal party election but not necessarily the general election. This is determined by the total votes all candidates from each list receive, which can include between 1 and 20 candidates. A party needs to get 1.5% of the total votes to move to the second round.

The candidates we chose to include were those who, at first glance, all consulting firms considered to have a chance to move to the second round. They are:

Lista “Juntos Por el Cambio”

• Patricia Bullrich
• Horacio Rodriguez Larreta

Lista “Unión por la Patria”

• Sergio Massa
• Juan Grabois

Lista “La Libertad Avanza”

• Javier Milei

Lista “Frente de Izquierda y los Trabajadores”

• Myriam Bregman
• Gabriel Solano

Lista “Principios y Valores”

• Guillermo Moreno

Lista “Hacemos por Nuestro País”

• Juan Schiaretti

The Context for Our Experiment

The context of our experiment is as follows: Most consulting firms identify Sergio Massa as the individual winner, with a percentage ranging from 30 to 33%, followed by Patricia Bullrich or Horacio Rodriguez Larreta with percentages that fluctuate between 27 and 31% for the winner. Next is Javier Milei with percentages ranging from 12 to 27%, Grabois with percentages between 5 and 7%, and the remaining candidates with percentages between 1 and 3%. These are the predictions for each candidate.

In terms of political parties, the same consultants stipulate a scenario in which “Juntos por el Cambio” would obtain between 32 and 38%, “Unión por la Patria” between 30 and 35%, “La Libertad Avanza” between 18 and 22%, and the rest of the parties below 3%.

These figures are quite notable. Although there are reasons for differences in the numbers, it’s logical to wonder why the votes by party are not relatively similar to the sum of votes for each candidate. This is a question without an answer, but one to which we will give importance.

The Elements of Our Experiment

As we have already said, to carry out our experiment, we used a simple (though tedious) data collection process. It was simply a spreadsheet where a value of 1 was assigned for each mention of the candidate made only in the comments.

Given that time was very scarce, we already had a factorial design system prepared in Python and Psycho-Py, with an artificial intelligence engine that analyzes the data through a regression analysis. This takes various variables for regression and compares them. One of these comparison variables is the result of an ANOVA (Analysis of Variance), while others were age (when possible), geolocation (likewise), proximity to the election, mentions in the media, and finally a qualitatively assigned value. This value, which we will call ϵ (greek letter Epsillon), is a variable that fluctuates between 0.1 and 1. It is determined by a mix of context, national history, economy, electoral history, and, honestly, “gut feeling.”

Getting results

As we all know, user behavior prediction is not an exact science. It’s true that Quantum UX makes it easier to predict in real time, but not necessarily in the long run. Furthermore: that’s exactly the advantage (and disadvantage) of Quantum UX!

So in this case, this is the raw data we extracted from our research before using the AI engine.

Please note that the data presented here is not the final data set. Instead, it is the preliminary information that will be fed into our AI engine for analysis. You might notice something that appears a bit unexpected: Javier Milei is the candidate with the most votes. This result is likely a statistical aberration, which is why we will employ our mathematical model to further investigate the phenomenon.

Applying AI and Quantum UX to User Behavior

So after the previous data, we ran our model and much to our surprise, the supposed statistial aberration not only didn’t disappear: it became even more prominent:

Once again, Javier Milei takes first place. Although this is not enough to secure first place for his party, he manages to surpass Sergio Massa, the candidate who everyone considered to be the individual winner.

In terms of the parties, there are not many surprises. Milei’s individual performance certainly alters overall predictions, but the final results still indicate that the winner will be Juntos por el Cambio, followed by Unión por la Patria, and then La Libertad Avanza.

Nevertheless, in our analytical model, we observed a rather unexpected technical tie between three distinct political parties. This outcome, while intriguing, is something we had not anticipated.

In all honesty, we find it challenging to believe that this scenario will actually materialize. Considering Milei’s recent performance in the media and his distinct anti-establishment rhetoric, it might appear that he has reached a peak in momentum. However, we assert that this is likely not sufficient to put him on equal footing with the two dominant coalitions that have historically polarized each election (analogous to the perennial battle between the U.S. Democrats and Republicans).

This theoretical tie has nonetheless provided an insightful and engaging perspective, one that may challenge conventional political thinking. While the hypothesis may seem far-fetched, it serves as a compelling exercise in exploring potential surprises in political landscapes.

As an intellectual experiment, this exercise on user behavior prediction has been quite stimulating and enjoyable. With the elections just six days away, the anticipation continues to build. Our team is eagerly awaiting the results, ready to delve into the data and provide timely insights. Rest assured, we will share our comments and analysis as soon as we have the necessary information at hand!

This experience reminds us that politics can be unpredictable, and even the most unlikely scenarios may provide valuable insights. The forthcoming elections promise to be an exciting time, and we are fully engaged in tracking the developments as they unfold.

Fabio Devin

Fabio is Bachelor in Marketing, Master in Human Computer Interaction and a UX Evangelist.

He worked for some of the biggest companies in the world. As such, he has developed entire systems and businesses for clients of all sizes.