This section will focus on the New York Airbnb and vacation rental market dataset. “With a total of 10,752 active Airbnb listings in the past 12 months, New York presents a significant short-term rental landscape. The key financial indicators for Airbnb and STR properties are promising, with an average nightly rate of $204 and a robust occupancy rate of 45.9%. These Airbnb rental metrics contribute to an estimated average annual revenue of $25,461 per Airbnb listing, highlighting the vacation rental investment potential within the United States region.”

We will answer 6 research questions:

1. Decompose performance into market-month demand and listing/host skill.

2. Forecast next-month (≈ next 30 days) occupancy using pacing/booking signals.

3. Simulate alternative pricing policies (counterfactual, observational).

4. Infer implicit objectives (ADR vs Occupancy vs RevPAR strategy clusters).

5. Detect demand shocks from market time series and identify winners/losers.

6. Detect anomalies (suspicious/strategic patterns) with unsupervised ML.

Question 1: Can we separate “market demand” from “host skill” in explaining performance?

In other words, would it be possible to build a model that compares market-week effects(demand) and listing effects(host/property skill)

To do this, we will analyze log_revenue ~ month + listing + controls and extract the month and listing coefficients.

We can use this data to create a simple graph, however, this doesn’t look that great. It’s easy to understand, but we can create a better looking one that’s more engaging. We will do that later.

Listing skill scores (after controls + market effects).

Now we attempt to predict the next month’s occupancy using pacing signals. This sets a target at the next occupancy, using previous pricing(variable on rate_avg) and pacing(booking_lead_time_avg, length_of_stay_avg) and reviews/last-month performance in a time-based split to make a prediction. We can then plot this prediction vs the actual occupancy data as follows:

Next, we can also define a price position per listing-month that is relative to similar listings actually found in the data. These can be grouped by three different price buckets(from low, to mid, to high). The outcome should return the next-month occupancy / revenue.

We can then compare outcomes across different matched groups. Thus this part of the project this isn’t a randomized experiment, so it has to be presented as a simulation that has limitations on how it will predict real life.

Now, we may cluster listings into different strategy archetypes. We will do this by analyzing the long-run averages (using variables such as booked_rate_avg and revenue/nights_total)

Finally, we can analyze the data by using some things learned from AP Statistics in high school: trying to detect shocks using z-score of month-over-month changes. Later, we could also try to add ruptures change-point detection, which is a Python library that finds points in a time series where behavior shifts. This could be like a sudden jump or drop in average occupancy, or a new trend in Airbnb costs. This would help for detecting changes in mean levels, trends, and variance.

Finally, we want a result that looks flashy, so we should determine who the winner or loser is around a shock month and compare pre vs post by categories.

For context, a shock month is a month where the market behaves abnormally.

• So like a sudden jump/drop in demand or pricing power.

A winner in our definition is a category of listings that benefits more(or declines less) around this shock than others.

We can do this by first picking a shock month t, and defining two windows

1. Pre: month t — 1 (or average of t — 2, t — 1)
2. Post: month t + 1 (or average of t + 1, t + 2)

Then we choose the metric that we want to care about, in this case we will use revpar, which is revenue / (vacant_days + reserved_dats)

We will find this for each listing category, post and pre and find the change by subtracting pre from post. Then compare this change for the listing to the overall market change.

If the change for the listing is greater than the overall market change, then it outperformed the market change and is a winner.

If it did not surpass the overall market change (delta market — delta listing < 0) , then it underperformed and is a loser.

This heatmap reveals how the calendar shapes demand and how much occupancy rates diverge across price tiers. Lower‑priced listings below $200 maintain relatively stable occupancy throughout the year. This suggests that travelers looking for budget-friendly AirBnbs, such as students, solo visitors, and long‑stay guests, have consistent demand regardless of the season.

As prices rise, demand begins to show more noticeable changes across the calendar. Listings around $200 to $600 show rises in occupancy around spring and summertime, from February through August. This pattern aligns with typical travel windows around school breaks, holidays, and warmer weather.

For listings around $600 to $900, this contrast becomes even more clear. Demand rises during the summer months, but unlike the lower tiers, these listings also see more activity in November and December. Travelers in this bracket likely include families and groups gathering for end‑of‑year events, as well as travelers willing to spend more for Thanksgiving and winter holidays.

The listings above $900 show the most extreme differences in occupancy across the year, and the highest occupancy rates in this tier are around double the occupancy rates of lower-priced listings. Demand spikes during peak summer months when high‑income travelers are most active, indicating that the demand for more luxury AirBnbs seems tied to periods of time associated with touring and traveling.

Across the differently-priced listings, it is clear that the higher the price of the listing is, the more strongly demand for AirBnbs rises and falls throughout the year. Cheaper listings remain steady year‑round while the fluctuating demand for more expensive listings reflects the timing of Los Angeles’ peak travel months.

Imagine yourself as an investor looking at the Los Angeles Airbnb market. The two options you have are a modest $120 per night apartment and a sleek $500 per night luxury loft in West Hollywood. Intuitively, you would think that the higher priced property must be a safer bet. Premium pricing signals quality, exclusivity, and affluent demand. But does higher price actually mean lower risk? To test out this assumption, I analyzed past Airbnb calendar data in Los Angeles. Properties are grouped into low, mid, and high tiers based on their average nightly rate. Instead of comparing average revenue, I measured revenue volatility. This is the calculation of the standard deviation of monthly revenue for each listing, capturing. how much income fluctuates over time. The results for this is shocking. Revenue volatility actually increased dramatically with price tier, with high priced listings exhibiting nearly five times the revenue volatility of low tier listings. This means that the most expensive properties experienced the largest swings in income. At this point you may be wondering why this might happen, and there could be a couple of different reasons. Luxury listings are often dependent on holidays, events, or summer time. When they are not busy, they lose bookings very fast. This is why their income is unstable and goes up and down a lot. On the other hand, cheaper listings attract more people.They are booked more consistently, so their income is also relatively more stable.

One appeal of AirBnB is its accessibility. Any ordinary homeowner can easily list their property on their own without the need for a real estate agent or such. However, this raises the question: does the way a home is listed on the AirBnB website affect its popularity? Specifically, how much do listing features affect the revenue of similar homes?

To answer this question, we created a regression model that predicts the expected revenue of a property solely from its location (latitude and longitude) and size (number of beds and baths). For simplicity, we only included listings that included the entire home rather than individual rooms. After calculating the expected revenue for certain homes, we subtracted this from the actual revenue for these homes, resulting in a residual for listing revenue that provided insight on how much the listing revenue varied based on certain listing features.

We compared the distribution of residuals for different listing features, including listing rating, number of photos, etc. For most, there was not a significant difference in revenue residual between different categories. The two features that displayed a noticeable difference between categories were cancellation policy and number of reviews (grouped by hundreds).

For almost all of the plots, however, the median residual appears to be less than zero. This means that the average expected revenue tends to be higher than the actual average revenue, implying that our regression model may be overpredicting revenue.

While the academic experience offered by UCLA is nuanced and diverse, the one common metric of student performance has always been letter grades. Grades provide insight into the performance of individual students, but they also provide insight into the performance of professors, courses, departments, and the institution as a whole. Being as central as they are to the student experience, UCLA’s grading is often a topic of discussion, with students trying to rationalize grades through common narratives like “grade inflation,” “weeder classes,” and professor dependency.

This project takes a comprehensive look at UCLA grade distributions over the past 15 quarters to better understand these patterns at scale. Using publicly available grading data from the UCLA Registrar’s Office, we address some of the common grading narratives discussed by students. We observe how grading has varied over time, by discipline, across quarters, and by professor. We also examine how professor ratings from the website Bruinwalk are related to their grade distributions. By looking at these patterns, this project aims to better understand the structure and variation of grading at UCLA.

Grades over time

To understand how grading has evolved with time, we plotted the average grades given each quarter starting with Fall 2021 and ending with Spring 2025. Average grades were computed by converting each quarter’s letter grades to their corresponding numerical GPA, then taking the average of all grades given in each respective quarter.

Overall, average grades remain fairly consistent throughout this time period, though there are minor fluctuations. Average grades for each quarter stay between 3.55 and 3.65, through they tended to be higher near Winter 2022 and started to dip around Winter 2023.

Something that surprised us was that there was a sharp increase in average grades between Winter 2024 and Spring 2024. During the Spring 2024 quarter, conflicts between police and student encampments in support of Palestine caused a portion of the quarter to move online. This potentially could have impacted student grades, considering that some exams might have been cancelled or taken online, or that professors may have been more lenient with grading in light of these circumstances. It is interesting that this was not the case during Winter 2025, when the first two weeks of the quarter were similarly moved online due to the outbreak of several large wildfires near the school’s campus.

Another trend we wanted to examine was the difference in average grades between STEM and humanities classes. A common point of contention between students is the debate over whether STEM classes are significantly more difficult than humanities classes, so we decided to see if this difference was empirically supported.

As shown in the graph, average grades for STEM classes were consistently and significantly lower than those of humanities classes. In the entire period from Fall 2021 to Spring 2025, the average grades across humanities classes never dipped below 3.6, while the average grades across STEM classes never spiked above 3.6. While grades are not an all-determining measure of class difficulty, the data definitely supports the notion that it is, on average, more difficult to achieve high grades in STEM courses than in humanities courses.

On Quarter Vs. Off Quarter

UCLA offers CS 31 (Introduction to Computer Science I), 32 (Introduction to Computer Science II), 33 (Introduction to Computer Organization), and 35L (Software Construction), which form the core introductory sequence for Computer Science majors. CS 31 covers an introduction to C++ programming, CS 32 covers object-oriented programming, CS 33 covers computer architecture, and 35L covers software construction. These courses are known amongst the student body for the high content difficulty, long project specifications, and fast-paced lecture style.

Though these are the critical foundation courses for CS majors, other majors and minors often require the completion of the CS 31 and 32 courses. For instance, students studying Computational & Systems Biology and wish to pursue the Bioinformatics Track need to complete both CS 31 and 32. Likewise, Electrical Engineering majors must also take the CS 31 and 32 courses for their undergraduate course requirements. Also, the Data Science Engineering minorAt UCLA requires the completion of both CS 31 and 32 to apply to the minor. Since the Statistics & Data Science major is known to have less course requirements, many students declare a minor, a popular pick being the Data Science Engineering minor.

As non-CS majors plan when to enroll in CS 31 and 32, there are common considerations amongst students. Enrolling in CS 31 during the Fall quarter is popularly known amongst the student body as the CS “on-quarter”. This is when most CS major students take the course, and it’s claimed to be more difficult as your grade would be in competition with pure CS majors. However, to the relief of non-CS majors, there is an option to take CS 31 in Spring quarter which is known as the CS 31 “off-quarter”. This is when most non-CS major students decide to enroll, as the course pace is said to be more adaptive to non-CS majors.

On top of all the discussion amongst students who need to take CS 31 and 32, an important consideration should be made — how do the grade distributions in “on” versus “off” quarters differ? We take a look into the grading distributions for Professor Smallberg’s CS 31 Fall (“on”) and Winter (“off”) lectures.

Professor Smallberg is a frequent lecturer for Computer Science at UCLA, especially CS 31 and 32. He is the only CS professor who teaches both CS 31 and 32 for both “on” and “off” quarters, which allows for a direct comparison between “on” versus “off” quarter types. Shown in the figure above, these faceted graphs are meant to compare his grading distributions for both CS 31 Fall (“on”) and Winter (“off”) quarters over the past four school years.

Each of the grade distributions for Professor Smallberg’s “on” and “off” quarters show similar trends and right-skewed shapes, but there are mostly higher proportions of CS 31 students in the “on” quarter that receive grades in the A range, as compared to students in the “off” quarter. This distinction is the most obvious for the 2023–2024 and 2024–2025 school years. For specific grades received, the proportion of students receiving As has decreased over the years, and consistently higher proportions of “off” quarter students receive a failing grade of F for each year. However, even as we note these deviations in proportions for Professor Smallberg’s courses, the differences are minimal. Rather than focusing on grade outcomes when deciding which quarter to take CS 31 in, one should consider other factors like different professors and the backgrounds of their classmates whom they may study with.

As we’ve seen differences in the grading distributions of CS 31 by analyzing the grade distributions of Professor Smallberg’s classes, it is also important to consider potential differences between professors. Besides Professor Smallberg, Professor Huang and Professor Stahl are also both frequent lecturers for CS 31. As lecture styles, topic focuses, and exam difficulty can vary by the professor, we must consider for these deviations. As Professor Huang has also consistently taught in the CS 31 off quarter (winter) for the past four years, we compare his grading distributions to those of Professor Smallberg’s CS 31 on quarter (fall).

While some professors might use a 97% cutoff for an A+, it is not generally expected in Professor Smallberg’s classes. In fact, in the past four years Professor Smallberg has not given out a single A+ in his CS 31 classes. Meanwhile, in the past three school years, around 20% of students in Professor Huang’s CS 31 course were able to receive an A+. From the graphs above, there is a higher number of students who receive grades in the A range for Professor Huang than for Professor Smallberg. Interestingly, after the initial differences between the proportions of students receiving an A- or higher, both professors’ grading distributions for grades B+ and lower are remarkably similar across the past four years.

Based on this quantitative data, if it is a student’s priority to perform well in their CS 31 course, it seems more likely to achieve this in Professor Huang’s CS 31 class in the off quarter compared to Professor Smallberg’s in the on quarter. However, what our analysis does not account for is student testimony and qualitative data of non-numeric perspectives. This is an important distinction to make, as our recommendations for which CS 31 professor to take are taken only from quantitative data.

How consistent are grade distributions among different professors who teach the same class?

UCLA offers STATS 100A (Introduction to Probability) and MATH 170E (Introduction to Probability and Statistics 1) as equivalent courses that undergraduate students can take to satisfy their major requirements. Choosing between equivalent classes and different professors taking those classes can be a challenging decision for many students. Students often assume that class content and syllabus are the main indicators of class performance. However, on analyzing data from these two classes, as well as among different instructors for each class, we see that professors are more significant predictors for the final letter grade that students receive. The visualizations also highlight differences between the two equivalent classes in terms of students’ final grades.

From the data visualizations for STATS 100A and Math 170E, we see that grade distributions are not consistent among different professors and vary immensely in both grade proportion and type. In the STATS 100A visualization, most professors give a relatively larger proportion of A/A- grades, for example, Professor Yingnian Wu, followed by B+/B/B- grades. On the other hand, some professors lean towards a grading scale that balances the number of A, B, and C grades given to students, for example, Professor Paik Schoenberg and Professor Montufar Cuartas.

Comparing equivalent classes STATS 100A and MATH 170E, it is seen that a high number of D/D-/F grades are given across different MATH 170E classes, compared to STATS 100A. This proportion varies across different MATH 170E instructors, but is noticeably higher than STATS 100A’s proportion of D/D-/F grades.

These differences between equivalent classes, and among the same class taught by different instructors, indicate that class and exam difficulty, teaching style, and instructor grading pattern or curve are highly variable. Although the overall course content and topics stay the same, final grades can look different for a student depending on the instructor or class equivalent that they decide to take. While the grade distributions do not provide detailed information regarding the difficulty of class content, they can provide students with important information while choosing classes and an overall picture of how grades typically vary with an instructor.

Choosing the right professor can make or break a student’s experience in the classroom, yet students at UCLA must rely on incomplete or anecdotal information when making those decisions. Bruinwalk, UCLA’s student review platform, offers one lens into teaching quality, but student ratings alone don’t tell the full story. The following question emerges: do professors who grade more generously earn higher student ratings, or do genuinely strong teachers rise to the top regardless of how they grade? Understanding this relationship will prove valuable for academic integrity, the validity of student evaluations, and how students should actually use these tools when registering for classes.

To explore this, we analyzed faculty statistics from the 2024–2025 academic year, mapping each professor’s Bruinwalk overall rating against their grade distribution score, while also accounting for the number of quarters and years they’ve been teaching. The heatmap visualization reveals a weak positive relationship between generous grading and higher student ratings; professors with higher grade distributions do tend to cluster toward slightly better Bruinwalk scores, but the correlation is very weak and inconsistent. For example, Xiaowu Dai is the top rated professor, with a perfect Bruinwalk score of 5.0, has a lower average GPA than Guang Cheng, a 3.2 to his 3.5.

The compelling takeaway is that teaching quality and style appear to matter far more than grades in determining how students rate a professor. Professors with decades of experience and many quarters of teaching tend to show more stable, reliable ratings; while outliers (both unusually high and low ratings) are more common among instructors with limited data points, meaning fewer quarters or years in the dataset. For students, this suggests that Bruinwalk ratings are a reasonably meaningful signal of teaching effectiveness, not simply a reflection of who gives out the most A’s.

• How much does overall bruinwalk rating compare to avg grade in class?
• How well can bruinwalk metrics predict avg grade in class?

Overall, our analysis shows that UCLA grading is influenced by multiple factors, including course content, professor, and academic discipline, with notable patterns emerging over time. While some commonly held narratives, like STEM courses being more challenging than humanities courses, are supported by the data, other assumptions, such as large differences between “on” and “off” quarters, appear less pronounced.

Differences between professors teaching the same course can be significant, highlighting the role of teaching style, grading patterns, and individual instructor decisions in shaping outcomes. At the same time, student perception from platforms like Bruinwalk provide some insight into teaching quality, but they are only weakly correlated with grades. This suggests that while students may use ratings to guide course selection, actual performance is influenced by more nuanced factors than anecdotal reviews alone.

By examining UCLA’s grade distributions at scale, this project provides a clearer, data-driven picture of how grading functions across courses, instructors, and departments. Understanding these patterns can help students make more informed decisions when planning their schedules, choosing professors, or navigating academic requirements, while also offering a broader view of how grading reflects the university’s academic structure and standards.

Music is possibly the most ubiquitous artform in human history. Whether you’re walking through the streets of medieval Europe, listening to the fantastical folk tunes coming from a nearby fiddleist, or studying in a coffee shop trying your hardest to ignore the baristas’ esoteric playlist, music is seemingly indispensable and everpresent. Perhaps more so than everpresent, music is distinct. No one person will have the same taste as another, just as no one person is the same as another. That being said, it is not unreasonable to be able to surmise one’s music taste from their outward appearance. Guessing the guy reading feminist literature at a coffee shop listens to Clairo is not exactly a shot in the dark. And yet, everybody knows that one person whose music taste and personality have seemingly no correlation. Whether they’re quiet and solemn, bumping “Party In The USA” on their headphones, or a jubilant ray of sunshine listening to Elliott Smith, it remains an intriguing phenomenon.

These scenarios beg the question, how accurate can we get? Is an angry person really more likely to listen to heavy metal? Does your melancholic friend actually like sad music? Inquiries like these were the genesis for our project, where we set out to answer the burning question: What does your music taste really say about you?

We dive into answering this with data pertaining to the Big Five Personality (BFI) traits and music features. The first main dataset is an experiment with 279 participants that collects personality profiles of participants and audio features of songs. The second contains information about 90,000 songs and their sentiments.

How does one’s favorite genre resemble how much they enjoy certain emotional tones?

Beginning with a broader understanding of what music genres can reveal, we examined several of the most popular genres and their valences. Valence in music is considered the pleasantness of a stimulus. A higher valence is happier and more uplifting, and a lower valence tends to be more somber. The histograms above show the distributions of valence across thousands of songs within each genre, which gives insight into how listeners’ favorite genres of music can be related to the mood of the song. Since valence corresponds to the emotional positivity of a song, the differences in these distributions suggest that genre preference may reflect a listener’s preferred emotional tone.

Pop music is highly clustered in the high valence range, indicating that overall, these songs tend to be uplifting. Listeners who prefer pop music may gravitate towards these feelings. Rock music, on the other hand, shows a much larger valence range. This wider spread implies people whose favorite genre is rock may be comfortable with both uplifting and slightly darker emotional tones — they may appreciate the variety. Electronic and indie music both have moderately high variation on average, but with a larger spread. These listeners may prefer the neutral emotional aspects of the music.

Overall, valence captures how positive or negative a song feels, so these genre-level patterns suggest that listening preferences may align with how individuals experience or regulate emotion. Some listeners appear to prefer consistently uplifting soundscapes, while others are drawn to a fuller emotional spectrum.

Does high extraversion imply a mainstream music taste?

We anticipated that individuals with higher Extraversion scores would prefer popular songs. So, we identified songs that were louder (high RMS), more energetic (activity), and more positive (high valence), as these features are strongly correlated with popular music tracks (found through correlations on Spotify Features Dataset). However, our analysis revealed near-zero correlations between Extraversion and these musical features, suggesting that extraverted individuals are not necessarily drawn to songs with characteristics associated with popularity.

Therefore, high extraversion cannot be conclusively correlated with enjoying songs that have musical features associated with high popularity. This raises the question — does that outgoing friend have a hidden music library that is more expansive than just the usual ten songs played at every party?

Do users with similar personality traits tend to prefer songs with similar audio/sound features?

It was surprising that extraversion didn’t seem to correlate with louder and more energetic music. In fact, this heatmap revealed quite the opposite.

Individuals with similar personality traits do exhibit predictable preferences for specific audio characteristics. Correlational analysis of the heatmap suggests several distinct patterns of association that support the hypothesis that personality acts as a predicting factor for musical taste. For example, our highest correlation coefficient of 0.13, indicates a positive association between extraversion and tender or sad music. This is quite fascinating as extraverts are usually seen as more outgoing and expected to listen to louder and more pop or hype music (proven false in the visualizations above). On the other hand, stereotypically, introverts are thought to listen to quiet and somber music — they keep to themselves and so does their music. Our lower correlation of -0.13 between agreeableness and RMS (how loud and aggressive a song is), instead shows that more agreeable users tend to favor music with lower overall sound.

While these correlation coefficients output to be moderate, the results confirm that personality traits are a crucial factor in shaping musical preferences.

Can we cluster users into groups based on music taste?

A natural next step in our journey, after realizing that personality traits correlate with musical preferences, was to determine if we could cluster listeners based on sonic values of favorite songs. To tackle this problem, we looked at the audio features (valence, activity, energy, etc) of songs listeners rated 4+/5, and plotted them using ‘kmeans’ and Principal Component Analysis (PCA). As seen in our visualization, there are three distinct clusters that appear based on the sonic values we measured: clusters 0, 1, and 2.

Through analyzing our principal components, we found that as one moves right on the x-axis (which represents activity and energy), positivity increases. This implies that the rightmost songs are lighthearted, bouncier, and more danceable. The y-axis was more balanced in emotional qualities, and represented a more general display of emotional expressiveness. Listeners towards the top of the graph may be more inclined to listen to emotionally complex songs, whereas those closer to the bottom may be less explorative in their listening.

Now that we have these clusters, what do they really tell us? To answer this, we created a series of box plots analyzing the clusters and various emotional traits of listeners, leaving out ones that are too similar for clarity.

We noticed that cluster 0 seemingly has more extroverted, sociable, productive individuals. Comparatively, cluster’s listeners are more anxiety driven, melancholic, and a little less productive. Cluster 2 is a nice middle ground — slightly less extroverted and social than cluster 0, but with a similar proclivity to anxiety and depression. Continuing with this chain of analysis, we veered towards the following heatmap, once again comparing clusters with their emotional traits, now emphasizing mean values.

This graph provides a clearer insight into the clusters. Taken together, these personality differences align with distinct patterns of music taste across clusters. Cluster 0, with higher scores in extraversion, sociability, and productivity, tends to align with happier, but lower energy music. Cluster 1, characterized by higher anxiety and lower productivity, is more associated with high energy, happier music. Cluster 2, occupies a middle ground favoring emotionally rich but balanced music that blends expressiveness with moderate energy, aligning with the listeners’ heightened compassion and respectfulness without extreme emotional volatility.

~

In conclusion, our analysis suggests that while music taste does not perfectly shape one’s personality, it does reveal meaningful patterns about how people experience and engage with personal feelings. Musical preferences often reflect the emotional environments individuals are drawn to, with some gravitating towards uplifting music and others preferring a wider emotion spectrum. Interestingly, several of our findings challenged common stereotypes about music and personality, with extraverts tending to prefer a more melancholic tone. Our clusters revealed that listeners can be grouped based on musical preferences, which reveal common emotional and behavioral tendencies amongst each group.

All in all, it’s safe to say that music taste functions more like an emotional reflection rather than a strict personality test. Our research highlights that our personalities do not need to be tied to a label like ‘Metalhead’ but rather to the specific sonic textures that we enjoy experiencing. Music is ultimately a medium to express oneself and while it cannot necessarily form the basis for who someone is, it offers important insights about how they process feelings, seek stimulation, or regulate their mood.

You finally got that first house. That first car. That first big purchase. It feels great; that feeling of luxury, satisfaction, and fulfillment fills your soul. Then it hits you. They hit you. The bills. The interest. That excitement you once felt soon turned to dread: the burden becomes too heavy to carry and the payments stop. Welcome to the dangerous territory of a loan default.

What exactly is a loan default? In its simplest terms, a default occurs when a borrower fails to uphold the legal obligations of a loan agreement. It isn’t just missing a single payment (called a delinquency); it’s the “point of no return” where the lender decides the agreement has been broken entirely, typically after 30 to 90 days of non-payment. It is a formal declaration that the trust between the bank and the borrower has collapsed, triggering a cascade of consequences from plummeted credit scores to the repossession of the very assets you once celebrated.

So, what causes a person to fall into this trap? Financial hardship is the most common culprit: unexpected job loss, medical emergencies, or a divorce can evaporate a person’s ability to pay overnight. However, as data often shows, the terms of the loan itself play a massive role.

Understanding these triggers is the first step in moving from financial vulnerability to financial resilience. The goal of our research is to accurately predict the probability of default through machine learning and the observation of default variables. To accomplish this, our team looked at a dataset from LendingClub, a company that provides loans, to find the most important variables in predicting loan defaults. In this article, we will dive deeper into the mechanics of default and how to navigate the waters before the ship begins to sink.

Let’s start with the basics: what variables did LendingClub make available to us, and what do they mean?

Putting the file into Visual Studio Code and retrieving just the columns outputs the variables that LendingClub collected from its clients. With 36 columns to look at, it’s important to differentiate the meaningful data with the junk. For example, some important variables to look at would be loan basics: the loan amount (3), funded amount (4), term (6), and interest rate (7). Moving on from that, we can see more information about the borrower: Do they own a home (10)? What’s their annual income (11)? What’s their debt to income (18)? Lastly, we need to know the outcome of their loan: did they repay it, or did they default (36)?

This gives us a basic idea of what we are looking for in our first visualization. We want to know what variable gives us the best chance of predicting a loan default. In other words, when someone defaults on a loan, which variable is the most present? Using a bit of machine learning, our computer was trained to find the variable that most accurately predicts loan defaults.

Our first visualization shows us the top ten most accurate predictors of loan default, from most important to less important. The visualization shows that the interest rate is the most important factor in predicting loan default, with an importance score of over 0.08. This makes sense because in the real world, higher interest rates are harder to pay off. Most people just see the small payments and forget to account for the added interest on each payment. Lastly, looking at the rest of the visualization, we can see that most of the variables present were the ones that we initially predicted when determining what was important for a loan default.

Next we have a heatmap that visualizes the loan default percentage across different loan purposes, helping identify which categories of loans are associated with higher repayment risk. It depicts a wide variability in default rate and emphasizes that small business and educational loans are more than twice as likely to default than car, wedding, and general major purchase loans.

One possible explanation for this is that the types of borrowers associated with these loans often differ in financial stability. Borrowers taking out small business or educational loans may face greater risk and delay in future income, often requiring larger loan amounts and longer repayment horizons which stack up. Alternatively, loans for cars, weddings, or other one-time purchases tend to be tied to more immediate and predictable financial situations, where borrowers may already have stable income allowing them to begin repayment immediately. This suggests that borrower characteristics and financial risk are a significant factor in predicting loan defaults.

Moving on, this graph is a combined visual of the default rates for two variables. The left hand side is the default rate for each number of inquiries and the right is the default rate for customers based on their debt to income ratio, based on lending club data. When identifying loan default factors, it is important to consider what is taken into account when performing risk management and when. Every time a client applies for credit, the lender records it as an inquiry. Inquiries happen after the loan is borrowed, while debt to income is analyzed by the lender before the loan is provided. While a debt to income ratio is a good indicator for a client defaulting a loan, inquiries were a greater indicator for which clients would default after the loan is provided. In terms of credit risk, this pattern is referred to as “credit hunger” where inquiries can be tied to applications for other loans, and in turn, higher future obligations.

The histogram above depicts the credit hunger effect, showing how each additional inquiry per client provides a significant increase in a risk of defaulting. Due to the fact inquiries tend to happen after the loan is provided, it cannot be used to determine which clients should be avoided. However, it tends to reflect a strong need for credit or a willingness to take on more debt to increase spending. Whether the application is accepted by the provider or not, the inquiry itself indicates a need for money, and thus the possibility of defaulting. While it cannot be used to assess the probability of defaulting before the loan is provided, the number of inquiries can be utilized by lending institutions to adequately prepare for possible defaults.

Lastly, this visualization is a box-and-whisker plot overlaid with a strip plot, built using a public Kaggle dataset, and designed to explore whether loan amount is a meaningful predictor of early default behavior. Each bin along the x-axis groups loans by size from 0K to 35K, while the y-axis tracks months until default, and the individual dots give us a raw, unfiltered look at how the data actually distributes beyond just the summary statistics.

What this chart makes pretty clear is that loan amount is not one of those strong predictors. Scanning across every bin, the box plots look almost identical, the medians sit in similar ranges, the interquartile boxes overlap heavily, and the whiskers stretch to comparable extremes across the board. If loan size genuinely drove default timing, you’d expect some kind of visible trend, like smaller loans defaulting faster or larger loans lasting longer, but that pattern just doesn’t show up here. That makes sense in the real world, whether someone borrowed $3,000 or $33,000, defaults tend to happen because of job loss, medical emergencies, or general financial instability, none of which are really tied to the original loan size. That said, the larger loan bins do show noticeably more variability, with wider spreads, longer whiskers, and more extreme outliers showing up in the 25K–35K range, suggesting that bigger loans introduce more unpredictability rather than a clear directional shift. So if anything, loan amount acts more like a noise variable than a signal variable, it widens the range of possible outcomes without reliably telling us which direction things will go.

Our research indicates that loan default is a multifaceted risk driven more by interest rates and loan purpose than by the initial loan amount. While machine learning identifies high interest rates as the most significant predictor of a “point of no return,” our visualizations also reveal that small business and educational loans carry twice the risk of automotive or wedding loans, largely due to the delayed income associated with those ventures. In the real world, these metrics are vital because they transform financial vulnerability from a mystery into a manageable set of data; identifying behaviors like “credit hunger” — where frequent inquiries signal a desperate need for more debt — allows both lenders and borrowers to spot a sinking ship before the trust between them collapses.

For those navigating these shark-infested waters, the best defense is a proactive strategy: prioritize low interest rates over the total loan size, maintain a healthy debt-to-income (DTI) ratio to absorb unexpected life shocks like medical emergencies, and avoid the trap of multiple credit inquiries that signal distress to the market. Ultimately, resilience comes from recognizing that while a loan amount creates noise and unpredictability, your specific financial habits and the terms of your agreement are the true signals of whether you will reach the safety of repayment or drift into the territory of default.

What really affects how people spend? Is it their age, the time of year or the type of products they choose to buy? At first, consumer behavior may seem predictable: during holidays all product categories see higher revenue, the people who buy more items spend more. This, however, is not always true. When you look closer, patterns are far less straightforward than they seem.

This analysis explains how spending changes based on different age groups, seasons, and product categories. It shows how consumer behavior is influenced more by life stages rather than simple, obvious trends. The data challenges common assumptions and reveals the complexity behind everyday spending.

Does spending increase linearly with age, or are there distinct age levels where product preferences change drastically?

When we look at average spending per transaction across different ages, there’s a small overall decline in spending as people get older, but the data also reveals some interesting fluctuations. For instance, spending tends to spike in the late teens to early 20s, dips slightly, and then rises again in the mid-to-late 30s. After that, there’s a noticeable drop in the early 40s, which is followed by smaller ups and downs in later years.

These changes urge us to think that life stages might play a bigger role in consumer behavior than just age alone. For example, early adulthood and mid-career years appear to overlap with higher spending, while other periods show more restraint. For marketers and businesses, this could mean that targeting strategies might work better when they focus on specific age brackets instead of assuming spending trends are uniform across a human lifespan. That is why understanding these complexities can give a more exact picture of how demographics can affect spending habits.

How does age purchasing behavior vary across seasonal cycles?

Analyzing the quarterly sales over categories shows distinct purchasing patterns among different ages. The data is segmented into 3 age groups (18–24, 35–36, 36+) to highlight how consumer spending habits shift with age throughout the year, offering retailers insights on how to better understand their target audiences.

Among the youngest age group, overall spending is the lowest out compared to the others and lacks a strong dominant category. Clothing starts out as the highest in Q1 but falls to the lowest in Q4, whereas Beauty and Electronics remain close in performance, floating around the 6–7 thousand dollar mark for total sales. The graph’s even distribution suggests that younger consumers have not yet established solid spending priorities and may be influenced by seasonal factors.

The 25–36 age group has higher total spending on average compared to the younger class. Beauty remains right under $10,000 for each quarter while Clothing is the highest total spent category in the first 3 quarters. The most notable trend in this segment is the trajectory of Electronics, which begins as the lowest-performing category but shows the strongest growth across the year, ultimately becoming the top category by Q4. Electronics’ great increase may reflect consumers’ increasing desire to invest in technology in these more settled stages of life, especially nearing the end of the year.

Lastly, the oldest age segment of individuals 36 and older are the highest and most consistent spending group across all ages. Electronics lead by a great margin throughout the entire year, with Clothing being second and Beauty being third. All categories here follow extremely similar seasonal trends, with increases in Q2, a slight dip in Q3, and a recovery in Q4. This age group possesses consistent and stable purchasing behaviors which suggests that older people have precise preferences and greater disposable income to sustain purchasing throughout the entire year.

What are some seasonal sales patterns and how do they impact total revenue?

When planning inventory and estimating revenue, it is essential to understand how different product categories behave over time. In order to reveal this information, it’s vital to dissect seasonal trends to understand which product categories drive overall revenue over certain quarters throughout the year.

While every category (Beauty, Clothing, and Electronics) show clear fluctuations throughout the year, changes in the sales of Electronics and Clothing show the strongest impacts on revenue. Specifically in May, October, and December, electronic sales spike significantly, which are periods most commonly associated with back-to-school purchases and holiday shopping. Notably, these peaks in Electronics sales coincide with the peaks in total revenue. During strong sales quarters, it is clear that Electronics acts as a primary revenue driver. Conversely, when sales in Electronics declines, such as during March and September, total revenue also declines. This shows how strongly the two correlate.

After Electronics, Clothing shows a moderate seasonality. Sales remain consistent, with a dip in early summer, and a rise again during the fall and early winter. One interesting observation is that when the sale of Electronics slightly dips, the sale of Clothing balances out this dip. Between June and July 2024, the sale of Electronics takes a slight hit, whereas the sale of Clothing increases. However, the total revenue seems to stay consistent, stabilising baselines.

Beauty, by contrast, shows the weakest seasonality, with sales staying the most consistent throughout all seasons. There are slight peaks during mid-summer and late fall, but nothing dramatic. As a result, Beauty does not contribute much to changes in the overall revenue curve, and acts rather as a reliable category throughout quarters.

Together, these patterns highlighted the revenue structures: Electronics drive major seasonal peaks; Clothing provides moderate variation; Beauty ensures consistency, and is the backbone salespersons can depend on. For businesses, this means that a

focus on Electronics during the seasons in which they are bought significantly more can maximize revenue opportunities, while also maintaining a strong backend focus on Clothing and Beauty to ensure steadiness.

Are higher transaction values associated with purchasing fewer high-priced items or more low-priced items?

Higher transaction values can stem from several possibilities: buying fewer expensive items, buying multiple low-priced items or a combination of both. But which factor contributes more? The graph reveals two clear patterns based on price level where the circle size is proportional to cost per unit. The orange cluster representing higher-priced items displays a steep upward trend. For example, purchasing just three to four high-priced items leads to transaction values above $1,000. The blue cluster represents lower-priced items that increase at a much slower rate. Remarkably, purchasing four to five low-priced items still results in a lower transaction value than buying a single high-priced item.

Why is this the case though? The much steeper slope of the high-price trend line shows that price per unit plays a far more influential role than quantity in driving higher transaction values. This likely occurs because transaction value is linearly

proportional to quantity but is multiplied by price. This causes higher-priced items to have a stronger impact even at low quantities while low-priced items require abnormally larger quantities to reach similar values.

Conclusion

These patterns show that consumer behavior is molded more than just simple trends like seasons or age. Life stages, pricing and product categories all interact to shape spending patterns. For businesses, this highlights the importance of moving beyond obvious assumptions and focusing on data-backed strategies.

Introduction

Has this ever happened to you? You come across a good flight deal, but when you come back to purchase a few days later, you find out that the price has almost doubled. You are not alone! According to the Federal Aviation Administration, about 3 million passengers fly daily on U.S. airlines, and a large number of these people might be doing the same thing as you, wondering when to book their ideal flights.

This frustrating problem sparked our curiosity: Is there actually a pattern behind these flight price changes, or are they just changing randomly? To find out, we analyzed a dataset of domestic flight prices collected across various search dates for the same departure dates, covering routes involving six major U.S. airports: ORD, LAX, JFK, LAS, BOS, and SFO.

An EDA Overview of Key Price Drivers

So, what is causing all these price fluctuations? To get a general sense, we investigate the key factors that shape ticket prices, from seasonality, cabin class, stop count, to days remaining before departure.

Firstly, we examine the season during which travelers fly. The travel patterns for passengers are not uniform throughout the year. Thus, we use Kernel Density Estimation to visualize the price distributions across summer, fall, and winter. As observed in Figure 1, summer ticket prices are more right-skewed, implying a higher average price. In comparison, fall and winter prices peak more sharply around $200–300, showing more concentrated and affordable fares. This aligns with our common sense that tickets during summer are more expensive due to high demand for vacations. However, within a particular season, ticket prices still vary greatly, suggesting that other factors also play significant roles in pricing patterns.

Other than seasonality, we analyze how different cabin classes (Economy, Premium Economy, Business, and First) shape price distributions. As shown in Figure 2, Economy class has the lowest median price. However, its numerous extreme outliers reveal that dynamic pricing may push fares far beyond the typical range when demand is high, sometimes even higher than those of Business class. Surprisingly, Premium Economy has the most stable and least deviated price range compared to the other three classes. This is probably because these types of tickets are set as a stable upgrade option for customers to get premium service. Meanwhile, First and Business classes show more fluctuating patterns, with wide IQRs and multiple outliers. These observations show that though cabin class establishes a rough pricing bracket, it cannot fully account for price variance.

Not all variables influence pricing as apparently as the cases above. Before looking at how the number of stops impacts price, we considered a potential confounding effect: in domestic flights, connecting flights tend to be long-haul while non-stop flights are mostly short-haul. To reduce the influence of this uneven data distribution, we applied feature engineering to group flight duration into three categories: shorter than 3 hours, 3–8 hours, and more than 8 hours. Below is the heatmap controlled for both flight duration and stop count.

The result is quite counterintuitive. No matter how long the flight is, ticket prices generally rise as the number of stops increases, yet drop slightly when there are more than 3 stops. Most people usually assume connecting flights to be cheaper because of inconvenience, but the data suggests a different trend for domestic flights. To further verify this, we plotted median ticket price against stop number for the top 10 routes. With route controlled, the pattern becomes clearer: the relationship between stop count and price is not simply linear. Connecting flights with moderate stops add extra detours and push prices higher, but too many stops can actually produce discounted tickets, likely reflecting airlines lowering prices to sell seats on unpopular itineraries.

How Does Booking Time Affect Flight Prices?

Next, we visualize how flight ticket prices change depending on how many days remain before departure across different cabin classes. The graph shows the median ticket price using 7-day bins to smooth out short-term fluctuations. Separate lines represent Economy, Premium Economy, Business, and First class fares, making it easier to compare how prices evolve for each class over time.

For Economy tickets, a clear pattern appears. Prices stay relatively low and stable when flights are booked far in advance, roughly 90–120 days before departure. As the departure date gets closer, prices begin to increase gradually. The rise becomes more noticeable within the final three weeks, with a sharp spike in the last few days before departure. This suggests that Economy fares are strongly affected by booking timing, with airlines charging more for last-minute purchases.

Business class shows a slightly different pattern. Prices are generally higher than Economy and start out relatively expensive even when booked far in advance. Interestingly, Business fares decrease somewhat between about 120 and 80 days before departure, which could reflect temporary discounts or promotional pricing. However, similar to Economy, prices climb again as the flight date approaches, especially within the last two weeks.

Premium Economy sits between Economy and Business in terms of price. Its prices gradually increase as departure gets closer, although the trend is not as smooth or pronounced due to smaller samples sizes.

First class fares remain consistently high across all booking windows. Prices decrease slightly between about 120 and 30 days before departure, but then begin to rise again in the final weeks, with some variation right before departure. Compared to the other classes, the overall trend is flatter, suggesting that First class prices are less sensitive to booking timing and more tied to their premium nature.

Predictive Modeling: When to Purchase the Ticket?

While our exploratory analysis highlights important patterns in flight pricing, these insights alone do not directly answer the key question travelers care about: when should you buy your ticket? To address this, we build a predictive model that translates observed patterns into actionable decisions.

We reframe the problem as a classification task. Specifically, we ask: will the price decrease in the next time step? This allows us to focus on directional movement, which is more stable and practical for decision-making.

The figure above illustrates how the likelihood of a price drop varies with time before departure. Instead of behaving randomly, prices exhibit structured patterns over time, suggesting that predictive signals do exist.

To build the model, we engineer features that capture these dynamics. These include short-term momentum (recent price changes), rolling averages and volatility measures, and temporal variables such as days remaining until departure. For instance, comparing the current price to recent averages helps determine whether a ticket is relatively expensive or cheap within its local context.

The feature importance plot shows that variables related to recent price behavior, such as deviations from recent averages and short-term changes, play a significant role in the model’s predictions. This reinforces our earlier findings that flight prices exhibit short-term trends rather than purely random fluctuations.

We then train a gradient boosting model, which is well-suited for capturing nonlinear relationships in tabular data. The model outputs a probability that the price will drop, which we convert into a simple decision rule: if the probability is high, the traveler should wait; otherwise, they should consider buying.

In evaluation, the model achieves approximately 74% accuracy in predicting price direction. While not perfect, this performance suggests that flight prices are not entirely random and that meaningful predictive patterns exist. More importantly, the model reveals a practical insight: some routes exhibit relatively stable pricing, while others are significantly more volatile. Understanding this distinction allows travelers to make more informed decisions about whether to buy immediately or wait for a potential price drop.

Conclusion

Our analysis shows that flight price fluctuations are far from random, they largely follow predictable patterns shaped by various factors. Together, these findings translate into a few practical takeaways for travelers.

Book early if you can. For Economy class, the sweet spot is around 90–120 days before departure, when fares tend to be at their most stable and affordable. Business and First class travelers, however, benefit from booking around 20–25 days in advance, when promotional pricing is most likely to appear.

Just decided on a last-minute trip? Try booking around 10 days before departure! That’s roughly the point where prices are unlikely to drop further for Economy. For Business and First class, however, it may be worth waiting until 1–2 days before departure, as airlines sometimes cut down prices due to unsold seats.

Note that these findings reflect general trends in U.S. domestic flights. Though exceptions do exist, some individual itineraries may not always follow these patterns. For international routes and flight-specific predictions, we are planning to train on a larger dataset and build a prediction website.

Stay tuned for more!

Introduction

The American Psychological Association noted the average attention span in 2004 to be two minutes and 30 seconds. Today, it ranges anywhere from 8.25 to 47 seconds. That’s not even half what it used to be. As our society seemingly continues to become more and more “brainrotted,” it’s only reasonable to ask the question: What is scrolling doing to us, and how’s it affecting our lives?

Entertainment online has become so abundant and easily accessible that it’s actually becoming a concern. With so many options, we naturally fall into the endless loop of searching and scrolling on apps like Youtube, TikTok, Instagram, etc., for the next thing to fill our minds, and the thing is, it’s affecting us in more and more ways than we think. From our own mental health to how we sleep, there are a number of issues to look at, and one of the best ways to understand what’s really happening beneath is to analyze the data at hand.

We conducted an analysis on social media usage, sleep patterns, stress levels, and content consumption habits across thousands of users and multiple datasets. Examining relationships between screen time and mental health indicators, we hoped to uncover how and to what extent our habits with digital media are shaping our lives.

We first wanted to take a look at whether or not people actually finish what they watch online. This KDE chart represents how far through a video the average viewer watches, across 1,125 randomly sampled YouTube videos. On the x-axis is the completion rate of videos, calculated by dividing a video’s view duration by its video duration. On the y-axis is the relative probability of a video having a certain completion rate, where a higher value means more videos cluster around that completion rate.

What is most noticeable is the peak, showing where completion rates are most common.

From the chart, we see a roughly bell-shaped curve centered around 50%, meaning most users on YouTube only watch around half the length of the videos they view. There’s also a considerable amount of density around 0–20%, indicating people tend to click off videos before even watching them.

Important to note is that this is a population-level average, mixed with videos that are dropped early, as well as videos that most people watch completely. It is still clear that most videos are left uncompleted. This distribution of data is also relatively widespread, likely meaning that video type is important as well. Some genres may consistently hold attention, while others might not.

What are the key aspects that increase video engagement and virality? — Amaan Jethani

The million-dollar question that everyone wants answered is “what actually makes a video on social media go viral?” There is a deep psychological aspect to understanding the gaze of the masses. Since there are multiple dimensions to this query, including trends, the algorithm itself, and most significantly, viewer engagement. As such, to answer this question, we created a correlation heatmap of the randomly sampled YouTube dataset to analyze informative patterns in viewer interaction with content.

At a glance, viewer engagement (which is measured by the proportion of a video the viewer watches) is negatively associated with duration. This trend echoes throughout our current understanding of declining attention-spans and can clearly be identified in this data set as well. However, what is more interesting to observe is the weak relationship between engagement and views, or the number of subscribers, which suggests that popularity and the size of channels do not necessarily translate into the attention span of viewers.

An overlooked aspect of virality on social media is the comments. An interesting positive correlation emerges between the comment percentage (measured as comments per view) and duration, as well as a modest positive correlation with the like percentage, indicating that interaction with the content (or “interaction intensity”) is a dimension of analysis distinct from merely understanding view volume. Nevertheless, this interaction intensity is interestingly weakly related to engagement, which implies that videos can be highly interacted with even when they aren’t completely watched, or the reverse. This aspect is quite intriguing, as the higher interaction a video gets, the more the algorithm itself pushes the video to others’ screens even when they may not be interested, which could also play a role in the weak engagement relationship.

Do shorter YouTube videos perform better?

As previously mentioned, data shows that people only finish around half of the video they are watching. This led us to question whether shorter videos performed better than longer ones.

To explore this, we looked at the relationship between video length and total view count on YouTube. Surprisingly, our analysis showed substantial overlap of video duration among videos with a high number of views and a low number of views. The duration of the videos was not correlated with the number of views they got.

Video length alone did not appear to significantly influence overall success. These findings suggest that factors other than duration, such as video quality or category, may play a more critical role in determining a video’s performance.

Mental Health

We often hear the phrase “social media is bad for you”. Many parents restrict their children from downloading social media apps like Instagram because they are afraid that it will result in depressive tendencies. But is this all a myth, or is social media really correlated to worsening mental health?

To find this, we decided to analyze the correlation between social media screen time and factors that are linked to mental health: sleep quality, happiness index, and stress level. Insomnia has been notoriously related to depression, as shown by a Stanford study finding that sleep-deprived people are ten times more likely to be depressed; lower numbers on the happiness index lead to a greater chance of depression; and the National Institute of Health has found that chronic stress may result in anxious depression. Thus, by recognizing the data trends for these three factors, we can recognize whether or not social media is closely linked to mental health.

The best way to recognize a relationship between two variables is to plot them against one another and look at the general trend. Since there are more than two variables being analyzed, we decided to create a graph with several dimensions. Hence, this 3D scatter plot in Figure 2 plots screen time, sleep quality, and happiness index on the x-axis, y-axis, and z-axis respectively while the points are color-coded based on stress level. On the right of Figure 2, the legend displays that individuals with minimal stress appear as purple on the 3D graph while individuals with high stress levels appear in pink. An intermediary level of stress is identified by orange points.

From this 3D plot, we can see that as screen time increases, both sleep quality and happiness index decrease. This is why the points start off rather high but slowly start to climb down. In addition to this, as screen time climbs close and closer to ten hours, the points become increasingly pink, indicating extremely high levels of stress.

This shows that the hours spent on social media really is closely related to factors that affect mental health, in which there is decreasing sleep time and happiness yet an increase in stress level.

Conclusion

Our analysis concludes that the social media impact is more defined by interrelatedness (how we view and consume content and the total amount of time spent consuming content) than any single measure or number. The unexpected correlations we found (linkages between screen time and decreased sleep, stress, and happiness) were not surprising; some of the results around the relationship between video length and performance are much more complex than we expected.

So many aspects of ‘brainrot’ we see are not simply related to individual actions but are complex events, which cannot be quantified via individual actions. This is in keeping with our increasingly global concern about our attention spans; thus, this highlights why this issue cannot simply be explained by saying, “Just scroll less.”

Through this project, we hope to provide tools to aid individuals and society in becoming more intentional about what they consume daily digitally, providing individuals with the ability to intentionally create and live healthier lives.

Tired of your current music rotation and feel like you’ve outgrown your favorite artists? Try SoundScout, our music recommendation engine backed by a powerful machine learning model that analyzes artist similarity and tag patterns. Our project explores how one might navigate building an artist recommendation engine — techniques, strategies, and questions we wanted to explore with our data!

Stack

SoundScout is a full-stack music recommendation app built to handle over 300,000 artists. The frontend uses HTML, CSS, and JavaScript to provide a smooth experience with real-time autocomplete and dynamic updates, while the backend powers a scalable recommendation pipeline using TF-IDF, SVD, K-Means clustering, and cosine similarity to deliver ranked artist suggestions tailored to your taste.

How can we group artists based on their content tags?

We first asked how artists could be grouped based on their tags. After applying TF-IDF, each artist was represented as a high-dimensional vector, capturing the importance of different descriptive tags. We then used K-Means clustering to group artists with similar tag profiles and applied SVD to visualize these groupings.

In this plot, each point represents an artist and color represents cluster membership. Artists that are close together share similar tag profiles, forming clear groupings. This confirmed that clustering was capturing meaningful relationships and gave us a structured way to group artists in our model.

How many clusters are optimal for our dataset?

To answer this, we created an Elbow Plot, which shows how clustering performance improves as the number of clusters increases.

The graph shows a clear “elbow” at k = 8, where improvements begin to level off. This indicated that 8 clusters was the best balance between accuracy and simplicity, and we used this value in our final model.

What are the most frequently occurring content tags?

We then examined which tags appeared most frequently in the dataset to better understand what drives similarity between artists. We compared tag frequencies before and after filtering out non-music-related tags.

The filtered results show that a small number of meaningful tags dominate the dataset. This step improved our feature quality, ensuring that clustering and similarity calculations were based on actual musical characteristics rather than noise.

Prospects

We considered areas where SoundScout could be improved to make the system more robust and user-friendly. One key limitation is the dataset itself. We wish to expand it to include more artists, especially more recent and mainstream ones, would improve coverage and make recommendations more relevant to current listeners. In addition, incorporating more features could significantly enhance the system. For example, adding artist images would improve the user interface, while introducing filters (such as by genre or popularity level) would allow users to better control and refine their recommendations. We also identified the need for more intuitive features overall, making the system easier to use and interact with.

In Los Angeles County, the impact of heat is not felt equally among all the districts. Research related to thermal inequality reveals that socioeconomic factors play a significant role in drastic differences of temperature within the region. In the county, according to a Caltech study, “lower-income neighborhoods have hotter surface temperatures than higher-income neighborhoods… [with] differences can be up to 36 degrees Fahrenheit at noon on a summer day.” With higher temperatures, increased risks of heat-related illnesses also arise within the communities facing such a disparity.

This article dives deeper into the factors that contribute to heightened heat records of communities within Los Angeles County, such as landscaping density, work conditions, household income, access to cooling centers, and population density. Throughout this analysis, the metric Heat and Health Index is used for comparison, which refers to a CDC-developed risk assessment to identify communities that are most likely to experience negative health impacts from heat based on historical temperatures, heat-related illnesses, and social/environmental factors. This metric is referred to throughout the article in the form of a Heat and Health Burden Score, which measures the average of percentile ranked historical heat and health burden indicators.

How does density of landscaping influence temperature within areas of LA county related to Heat and Health Index Scores?

This visualization is a Kernel Density Estimate (KDE) and is used to show the relationship between the Heat Health Index and the percent coverage of tree canopy. The “density” shown on the vertical axis represents the probability density of the data points at a specific value. The “tree canopy coverage” percentage is weighted by the population within each of the zip codes representing an area in Los Angeles county.

Based on the chart, the plot shows that neighborhoods with the lowest tree canopy coverage (0–5%) face the highest risk of heat-related illnesses, peaking the highest at the end of the spectrum. Areas with moderate tree coverage (10–15%), on the other hand show a high probability of a lower HHI score. The 5–10% canopy group, though, does show that health risks are inconsistent in relation to canopy density, however it still demonstrates a lean towards a higher probability of heightened heat risks. It is worth mentioning that the 15–20% group does tend to have a probability of a higher heat-related illness risk.

This color-coded scatterplot represents how related the Heat-Health Index score of an area relates to the number of extreme heat days recorded within the area within Los Angeles County. The hue of the plot represents the weighted tree canopy coverage with the same distributions as that of the last visualization, with a deeper color representing a higher percentage of weighted tree canopy coverage. A grey-dotted non-linear regression model is included within the scatterplot to show the relationship between the two variables.

The model’s R² value of 0.994 indicates a strong, positive correlation between extreme heat and public health risks, while highlighting the protective role of urban greenery. Notably, areas with higher tree canopy coverage (10% or more) tend to cluster at the lower end of both the heat frequency and health risk scales, whereas areas with minimal shade are disproportionately represented in high-heat, high-risk scenarios.

How does household income/work conditions relate to neighborhood temperature?

This scatterplot correlates the relationship between median household income and the overall Heat-Health Index (HHI) score across 559 Los Angeles neighborhoods. In this visualization, each data point is colored to its specific Heat-Health Burden (HHB), which showcases the physical heat exposure, where deeper blue indicates extreme temperatures and environmental heat. The black dashed linear regression line models the trend between wealth and risk.

As income increases, the data points transition from a darker blue (high heat burden) at the top-left to light green (low heat burden) at the bottom-right. This suggests that neighborhood wealth is a primary predictor of physical heat exposure, with lower-income areas suffering the most environmental heat stress. The model’s R² value of 0.545 reveals a decently strong negative correlation, showcasing that 54% of the variance in a neighborhood’s heat risk is explained by median income.

To further explore socioeconomic trends with “heat gap,” this violin plot categorizes Los Angeles into four distinct income brackets, ranging from “Lowest Income” to the “Highest Income” to compare the distribution of risk scores. A higher heat risk score means more risk. The visualization utilizes a color palette from dark blue (high danger) to light green (low danger) to highlight the disparity in public health vulnerability. The grouping of individual data points is overlaid on the violins to show the spread in each type of neighborhood.

The results display inequality in environmental safety. Neighborhoods in the poorest 25% have an average heat risk score of 0.79, placing the majority of communities within the “Critical Risk Zone.” In contrast, the wealthiest 25% of neighborhoods have a lower average risk of 0.47. The violin shapes confirm this disparity: the wealthiest bracket is bottom-heavy, indicating a mass of low-risk scores, the poorest bracket is top-heavy, showing that lower-income residents have almost no access to low heat risk environments.

Are neighborhoods experiencing the highest levels of heat exposure also the ones with the greatest access to cooling centers in Los Angeles County?

This visualisation presents a binned scatterplot that examines the relationship between levels of heat vulnerability and access to cooling centres across Los Angeles ZIP codes. The ZIP codes were then grouped into 20 equal sized bins based on their Heat and Health Index score and a mean number of cooling centers (per 10,000 residents within the bins) was calculated- this is shown by the Y axis. The color gradient of the points represents the average level of vulnerability to heat within each ZIP code bin. The grey shaded area around the regression line represents the 95% confidence interval, illustrating the range within which the true relationship is likely to lie.

By interpreting the plot it is evident that there is a positive association between heat vulnerability and access to cooling infrastructure per ZIP code bin. As the Heat and HealthIndex (HHI) increases, the average number of cooling centres per 10,000 residents simultaneously increases. This suggests that there is concentration of cooling centers within areas that suffer greater heat-related health risks. However, there is substantial variation across ZIP codes as identified by the wide confidence interval surrounding the regression line. Whilst typically, communities more vulnerable to the heat-related health risk have better access to cooling centers per capita, the distribution is still uneven.

In summary, the visualisation indicates that whilst there is a positive association between cooling infrastructure allocation and community heat vulnerability, this may be inconsistent across the county as indicated by the strength of the relationship. Although access tends to increase with vulnerability, the overall per-capita provision of cooling centers remains low across much of the county, suggesting that expanding infrastructure may be necessary to adequately meet heat-related health risks.

How does population density relate to heat exposure?

This visualization is a bivariate choropleth map of Los Angeles ZIP Code Tabulation Areas (ZCTAs) that compares population density (people per square mile) with the Heat and Health Burden score (HHB Score). The bivariate legend uses four buckets for each variable. D1–D4 represent population density quartiles and H1–H4 represent HHB score quartiles, and 1 = lowest 25%, 2 = 25–50th percentile, 3 = 50–75th percentile, and 4 = highest 25%. Each zip code’s color reflects the combined category (e.g. H4-D4 is the combination of highest quartile density and highest quartile HHB score).

Based on this visualization, there does not appear to be a single consistent relationship between population density and HHB across Los Angeles zip codes. For example, both lower density areas (D1, D2) and higher density areas have higher HHB scores (H4).

To further explore the relationship between population density and HHB score, we can at the scatterplot of HHB score vs. population density (log scale) across Los Angeles ZCTAs.

This scatterplot compares population density (x-axis, shown on a log scale) with HHB score (y-axis) across Los Angeles ZCTAs, where each dot represents one ZCTA. Because population density varies widely across Los Angeles, the log scale on the x-axis helps spread out the x-axis by factors of 10 (e.g. 1 → 10 → 100 → 1,000 → 10,000) so the low-density areas aren’t compressed. The HHB score on the y-axis ranges from 0 to 1 and is based on the percentile of scores, so higher values indicate greater relative burden. The fitted trendline overlaid on the scatterplot points is a best-fit line of the pattern, and it shows a slight upward trend at the higher densities.

The fitted trendline suggests a very weak relationship between population density and HHB score. The small p-value (p = 0.0314) tells us the association is statistically significant, meaning that it is unlikely that the trend occurred purpley due to chance and there is evidence of some relationship. However, the low R2 value indicates that populational density is not a meaningful predictor of HHB differences across zip codes. In fact, population density only explains ~1.6% of the variation in HHB (R2 = 0.0158). This data suggests that other factors play a larger role in affecting HHB scores, so it is important to consider other variables such tree canopy coverage, landscaping density, income levels, and access to cooling infrastructure.

Conclusion

In our analysis, we observed that tree canopy coverage is shown to have a statistically significant correlation with lower HHI scores. Data reveals that tree coverage is associated with lower number of days with extreme temperatures recorded, although the benefits appear to be unequally distributed. Access to cooling centers also provide insights, as that while cooling centers are generally more common in high-risk areas, their availability is inconsistent. Factors of income reveal a noteworthy negative trendline between income and the Heat and Health Index score, demonstrating the significant relationship between socioeconomic factors and the HHI score.

In contrast, factors such as population density do not appear to have a significant impact in relation to the HHI score. This suggests that heat vulnerability and inequality within Los Angeles County is more of a socioeconomic equity issue that requires targeted infrastructure interventions rather than an environmental factor.

Introduction

Land vs Water. In sports, there is no bigger difference in terrain than between track and field and swimming. Both battle against the forces of nature — runners fight gravity and ground reaction forces, while swimmers fight drag and buoyancy. Yet for both, the goal remains the same: go as fast as you can.

At the elite level, physical differences become decisive factors: height becomes an advantage, weight translates to force, and age is a balance between peak athleticism and experience. Especially in these events, where outcomes are determined by hundredths of seconds, every small detail matters — whether that’s a slightly longer stride, a more efficient stroke, or a couple more years at the biggest stage.

Of course, the greatest stage for both sports remains the Olympics, a prestigious event held once every four years, where only the cream of the crop athletes from around the world get a chance to compete. Therefore, to medal in the Olympics is one of the greatest achievements attainable, requiring an athlete to be, quite literally, one of the best in the world. These athletes have honed their respective fields, both physically and mentally, mastering the unique physical demands of their disciplines.

Although both sports share similar objectives, the profile of a “fast” athlete varies between track and field and swimming. One could expect the ideal build for running, which requires generating force into the ground, to contrast with the ideal build for swimming, which minimizes drag through water. As such, this leads us to our question: at the Olympic level, how do attributes like height and weight, age, and prior Olympic experience relate to performance in track and field versus swimming?

Is Taller Better?

Conventional wisdom says that being taller is an advantage in sports: for basketball and volleyball, it’s almost self-explanatory, especially when professional athletes tower above the average human height.

But how does this translate for track and field and swimming, and is it more important for one discipline over the other?

Comparing Through World Record Times

Now obviously, running on land is faster than swimming. However, there is a trend, as shown in Table 1, in which world record track and field events seem to be a few seconds off from those of swimming events that are a quarter of the distance.

Table 1: For Track and Field, times are taken from outdoor tracks, which are what’s used in the Olympics. Similarly, swimming times were recorded in 50m (Olympic standard) pools, and are freestyle, which is the fastest of the four strokes.

Therefore, one way to compare swimming and track and field events is by world record times. This approach seeks to standardize duration, which, assuming events of similar duration have similar mental and physical demands, compares events based on intensity. Using this method, the height distributions of Olympic medalists were analyzed as they exemplify the peak physical demands to excel in their respective sports.

Analyzing the boxplots (Figure 1) of 50m swimming and 200m track and field medalists, the shortest sprint distance pair, a clear pattern emerges that 50m swimmers tend to be taller than 200m track and field athletes for both genders, as indicated by the median.

This suggests that being taller might play a more prominent role in swimming performance than in track and field running. One possible explanation is that taller swimmers benefit from a longer wingspan, allowing them to pull more water with each stroke. While taller runners might gain a similar advantage through longer stride length, enabling them to cover more ground with each step, the boxplot indicates that the association between height and performance is stronger in swimming than in track and field. This suggests that running relies less on absolute height, with other elements — like stride frequency and ground contact time — playing a greater role.

Looking at the whiskers, elite performance among women seems to be achievable across a bigger height range. This shows that women’s heights are less narrowly concentrated than men’s, and may suggest that height is a less restrictive characteristic in women’s events than in men’s.

In Figure 2, which models the height distributions for the 1500m running and 400m swimming event, both considered middle-distance events, the trend of taller swimmers than runners is followed.

However, as indicated by the whisker, the men’s 400m swimming event contains a greater range than the women’s 1500m track and field event, which does not follow our previous trend. It seems that at longer distances, the “ideal height” is less prominent as each boxplot has a greater height range, indicating that various heights found medaling success.

Comparing Through Distance

While the previous method compared events through duration, another approach is to keep the distance constant, matching 100m swimming with 100m track and field events. By standardizing distance, this approach highlights the effects of the different environments — land vs water.

Interestingly, the same trend of swimmers being taller than runners is present. It seems that even for events of different durations, which should have separate mental and physical intensities, the height advantage remains more prominent within swimmers.

The Golden Age: When do Olympians Peak?

A common belief in sports is that younger athletes are faster and therefore are more likely to win. The question is whether this holds at the Olympic level: does youth or years of seasoned experience provide a better competitive edge? Analyzing the age distribution of medalists in both swimming and track and field, a clear pattern emerges about when athletes are more likely to reach their competitive peak. In general, swimmers peak earlier than runners, but the specific “Golden Age” is often dictated by event distance.

The Kernel density plot (Figure 4) shows the probability density of winning a medal across ages, separated by gender, sport, and event distance. Rather than counting individual athletes, the curves highlight where medalists are more concentrated, allowing for simplicity in identifying the ages where Olympic success is most common.

To compare physiological peaks, events were categorized by intensity. Swim — Sprint events (50m, 100m), Swim — Distance events (400m, 800m, 1500m), Track — Sprint events (100m, 200m, 400m), and Track — Distance events (5000m, 10000m, marathon)

The Divide in Swimming vs. Running

Across both men’s and women’s events, swimming medalists are concentrated in the late teens to early twenties. This suggests that elite swimming performance typically occurs at an earlier stage in an athlete’s career. Several factors that contribute to this pattern include how sprint power, flexibility, and muscular performance tend to peak early. As a result, many swimmers reach peak form while still in college or even in high school.

In contrast, track medalists, particularly those of distance events, tend to be older in age (mid-to-late twenties). Distance events rely heavily on endurance, pacing strategy, and even aerobic capacity, all of which improve with experience and years of training. Unlike explosive sports, distance performance benefits from accumulated mileage, allowing elite runners to thrive well into their late twenties.

Therefore, it comes as no surprise that sprint events skew younger than distance events in both sports. Short races demand explosive speed and reaction time, while distance events reward efficiency and stamina. This distinction clarifies why even within the same sport, different events produce different age profiles for medalists.

Overall, men’s and women’s results follow a nearly identical trend: swimmers medal younger than runners, and distance athletes medal at an older age than sprinters. While the exact peak age varies for each Olympic sport, the overall progression from younger sprint specialists to older endurance athletes is constant.

Revisiting the “Younger is Better” Myth

These patterns demonstrate that the common perception about age and sports is indeed a misconception, at least on the Olympic level. Youth alone does not determine success at the Olympics level, and instead, peak performance is a reflection of the physiological and strategic aspects of each event. Events that emphasize explosive power tend to reward younger athletes, while those requiring endurance and tactical maturity favor the experienced.

Understanding the “golden age” ranges tells us that there is no single timeline for athletic success. For spectators, we can appreciate the diversity of athletic pathways that lead to Olympic success. For athletes, these patterns indicate that peak performance may arrive earlier or later based on training history and the physical demands of an event.

Conclusion

Through the lens of Olympic data, it is clear that the “ideal” blueprint for success extends beyond just the objective: go as fast as you can.

Our analysis of height distributions confirms that being taller is a more significant prerequisite for elite swimmers than for runners. The aquatic environment rewards longer limbs and a greater wingspan to overcome water resistance. In contrast, runners, who are still tall by general standards, exhibit a greater range of heights, implying that stride frequency or even ground force production can compensate for a shorter stature.

Furthermore, the presence of the “Golden Age” of performance is divided between each sport. Swimming tends to favor youth, where peak flexibility and explosive power allow athletes to make waves early into their athletic career. Track and field rewards maturity, particularly in distance events where aerobic endurance and tactical patience peaks in mid-to-late twenties.

Next time you watch the Olympics, look beyond the stopwatch and look at the athletes themselves. Whether it is the tall, young sprinter racing through the pool, or the seasoned runner battling through the final lap of a race, both represent the optimal archetype for their sport. The path to the podium is not universal, but a testament to the many ways that human potential can reach its peak.