Analysis of Top Selling Beats on Beatstars

Beatstars Top Chart Analysis: How to be a successful producer?

Intro

beatstars.com is the World’s #1 marketplace to buy & sell beats. Every day, producers who post their beats on this platform are looking for more and more ways to promote, as they are constantly in the huge competition of the global market. It became interesting for me to study the main features of the beats that are currently in the top-selling chart of the platform, so I decided to web-scrape Beatstars top-selling beats info and producers info.

The process of web scraping was done by RSelenium and was divided to stages: Stage 1 - processing the chart itself (“https://www.beatstars.com/top-charts”). In the result I receive a dataset that consists of beat title, producer name, place of beat in the chart and links to presented beats and producers Stage 2 - loop function through every beat page to collect beat info Stage 3 - loop function through every producer page to collect producer stats

As the result I have two datasets: “df_beats” contains information about beats (such stats as likes, reposts, plays and such factors as price, bpm, key, etc.) “df_producers” contains stats about producers whose beats are presented in the chart But still these datasets need Preprocessing to work with them.

So, the goal of that work is to receive some insights about factors of success on Beatstars. I would like to analyze the correlation and maybe even try to create a model of prediction of success.

Cleaning the Dataset

# Fix some errors
df_beats[364,1] = 362

# Deleting Track that been brought to chart due to advertisment
df_beats1 = subset(df_beats, place!="-")

# Change class of place
df_beats1$place = as.numeric(df_beats1$place)

# Remove technical variables
df_beats1 = dplyr::select(df_beats1,-beat_href)

# Change k to thousands and M to millions
df_beats1$beat_like_k = str_detect(df_beats1$beat_like,"k")
df_beats1$beat_like = str_remove_all(df_beats1$beat_like,"k") %>% as.numeric()
df_beats1$beat_like = ifelse(df_beats1$beat_like_k==T,df_beats1$beat_like*1000,df_beats1$beat_like)
df_beats1 = dplyr::select(df_beats1,-beat_like_k)

df_beats1$beat_repost_k = str_detect(df_beats1$beat_repost,"k")
df_beats1$beat_repost = str_remove_all(df_beats1$beat_repost,"k") %>% as.numeric()
df_beats1$beat_repost = ifelse(df_beats1$beat_repost_k==T,df_beats1$beat_repost*1000,df_beats1$beat_repost)
df_beats1 = dplyr::select(df_beats1,-beat_repost_k)

df_beats1$beat_plays_k = str_detect(df_beats1$beat_plays,"k")
df_beats1$beat_plays_m = str_detect(df_beats1$beat_plays,"M")
df_beats1$beat_plays = str_remove_all(df_beats1$beat_plays,"k")  
df_beats1$beat_plays = str_remove_all(df_beats1$beat_plays,"M") %>%  
  as.numeric()
df_beats1$beat_plays = ifelse(df_beats1$beat_plays_k==T,df_beats1$beat_plays*1000,
                              ifelse(df_beats1$beat_plays_m==T, df_beats1$beat_plays*1000000,df_beats1$beat_plays))
df_beats1 = dplyr::select(df_beats1,-beat_plays_k,-beat_plays_m)

# Change date format
library(lubridate)
df_beats1$beat_date=mdy(df_beats1$beat_date)

# Remove $ signs
df_beats1$license1_price = gsub('[$]','',df_beats1$license1_price) %>% as.numeric()
df_beats1$license2_price = gsub('[$]','',df_beats1$license2_price) %>% as.numeric()
df_beats1$license3_price = gsub('[$]','',df_beats1$license3_price) %>% as.numeric()
df_beats1$license4_price = gsub('[$]','',df_beats1$license4_price) %>% as.numeric()
df_beats1$license5_price = gsub('[$]','',df_beats1$license5_price) %>% as.numeric()
df_beats1$license6_price = gsub('[$]','',df_beats1$license6_price) %>% as.numeric()

# Additional features of naming
## Does beat_title include info about price
df_beats1$beat_title_price = str_detect(df_beats1$beat_title,"[$]")

## Does beat_title include info about special offer
df_beats1$beat_title_offer = str_detect(str_to_lower(df_beats1$beat_title),"buy")

## Does beat_title include info about type of beat
df_beats1$beat_title_type = str_detect(str_to_lower(df_beats1$beat_title),"type")

# License Prices
## MP3
df_beats1$mp3_1 = ifelse(is.na(df_beats1$license1_type)==T, 0, ifelse(df_beats1$license1_type=="MP3", 1, 0)) %>% as.numeric()
df_beats1$mp3_2 = ifelse(is.na(df_beats1$license2_type)==T, 0, ifelse(df_beats1$license2_type=="MP3", 1, 0)) %>% as.numeric()
df_beats1$mp3_3 = ifelse(is.na(df_beats1$license3_type)==T, 0, ifelse(df_beats1$license3_type=="MP3", 1, 0)) %>% as.numeric()
df_beats1$mp3_4 = ifelse(is.na(df_beats1$license4_type)==T, 0, ifelse(df_beats1$license4_type=="MP3", 1, 0)) %>% as.numeric()
df_beats1$mp3_5 = ifelse(is.na(df_beats1$license5_type)==T, 0, ifelse(df_beats1$license5_type=="MP3", 1, 0)) %>% as.numeric()
df_beats1$mp3_6 = ifelse(is.na(df_beats1$license6_type)==T, 0, ifelse(df_beats1$license6_type=="MP3", 1, 0)) %>% as.numeric()
df_beats1$isonemp3 = rowSums(df_beats1[,c("mp3_1", "mp3_2", "mp3_3", "mp3_4", "mp3_5", "mp3_6")])

df_beats1$MP3 = ifelse(df_beats1$isonemp3==0, NA, 
                       ifelse(df_beats1$mp3_1==1,df_beats1$license1_price,
                              ifelse(df_beats1$mp3_2==1,df_beats1$license2_price,
                                     ifelse(df_beats1$mp3_3==1,df_beats1$license3_price,
                                            ifelse(df_beats1$mp3_4==1,df_beats1$license4_price,
                                                   ifelse(df_beats1$mp3_5==1,df_beats1$license5_price,
                                                          df_beats1$license6_price)
                                                   )
                                            )
                                     )
                              )
                       )

## WAV
df_beats1$wav_1 = ifelse(is.na(df_beats1$license1_type)==T, 0, ifelse(df_beats1$license1_type=="MP3 AND WAV", 1, 0)) %>% as.numeric()
df_beats1$wav_2 = ifelse(is.na(df_beats1$license2_type)==T, 0, ifelse(df_beats1$license2_type=="MP3 AND WAV", 1, 0)) %>% as.numeric()
df_beats1$wav_3 = ifelse(is.na(df_beats1$license3_type)==T, 0, ifelse(df_beats1$license3_type=="MP3 AND WAV", 1, 0)) %>% as.numeric()
df_beats1$wav_4 = ifelse(is.na(df_beats1$license4_type)==T, 0, ifelse(df_beats1$license4_type=="MP3 AND WAV", 1, 0)) %>% as.numeric()
df_beats1$wav_5 = ifelse(is.na(df_beats1$license5_type)==T, 0, ifelse(df_beats1$license5_type=="MP3 AND WAV", 1, 0)) %>% as.numeric()
df_beats1$wav_6 = ifelse(is.na(df_beats1$license6_type)==T, 0, ifelse(df_beats1$license6_type=="MP3 AND WAV", 1, 0)) %>% as.numeric()
df_beats1$isonewav = rowSums(df_beats1[,c("wav_1", "wav_2", "wav_3", "wav_4", "wav_5", "wav_6")])

df_beats1$WAV = ifelse(df_beats1$isonewav==0, NA, 
                       ifelse(df_beats1$wav_1==1,df_beats1$license1_price,
                              ifelse(df_beats1$wav_2==1,df_beats1$license2_price,
                                     ifelse(df_beats1$wav_3==1,df_beats1$license3_price,
                                            ifelse(df_beats1$wav_4==1,df_beats1$license4_price,
                                                   ifelse(df_beats1$wav_5==1,df_beats1$license5_price,
                                                          df_beats1$license6_price)
                                                   )
                                            )
                                     )
                              )
                       )
df_beats1 = dplyr::select(df_beats1, -mp3_1, -mp3_2, -mp3_3, -mp3_4, -mp3_5, -mp3_6, -isonemp3,
                   -wav_1, -wav_2, -wav_3, -wav_4, -wav_5, -wav_6, -isonewav)
df_beats1 = dplyr::select(df_beats1, 
                   -license1_price, -license2_price, -license3_price, -license4_price, -license5_price, -license6_price,
                   -license1_type, -license2_type, -license3_type, -license4_type, -license5_type, -license6_type)

# Remove technical variables
df_producers1 = dplyr::select(df_producers, -unique.author_href.)

# Renaming of columns
colnames(df_producers1) <-  c("author_name", "author_followers", "author_plays", "author_n_tracks")

# Change k to thousands and M to millions
## author_followers
df_producers1$author_followers_k = str_detect(df_producers1$author_followers,"k")
df_producers1$author_followers_m = str_detect(df_producers1$author_followers,"M")
df_producers1$author_followers = str_remove_all(df_producers1$author_followers,"k")  
df_producers1$author_followers = str_remove_all(df_producers1$author_followers,"M") %>%  
  as.numeric()
df_producers1$author_followers = ifelse(df_producers1$author_followers_k==T,df_producers1$author_followers*1000,
                              ifelse(df_producers1$author_followers_m==T, df_producers1$author_followers*1000000,df_producers1$author_followers))
df_producers1 =dplyr:: select(df_producers1,-author_followers_k,-author_followers_m)

## author_plays
df_producers1$author_plays_k = str_detect(df_producers1$author_plays,"k")
df_producers1$author_plays_m = str_detect(df_producers1$author_plays,"M")
df_producers1$author_plays = str_remove_all(df_producers1$author_plays,"k")  
df_producers1$author_plays = str_remove_all(df_producers1$author_plays,"M") %>%  
  as.numeric()
df_producers1$author_plays = ifelse(df_producers1$author_plays_k==T,df_producers1$author_plays*1000,
                              ifelse(df_producers1$author_plays_m==T, df_producers1$author_plays*1000000,df_producers1$author_plays))
df_producers1 = dplyr::select(df_producers1,-author_plays_k,-author_plays_m)

## author_n_tracks
df_producers1$author_n_tracks_k = str_detect(df_producers1$author_n_tracks,"k")
df_producers1$author_n_tracks_m = str_detect(df_producers1$author_n_tracks,"M")
df_producers1$author_n_tracks = str_remove_all(df_producers1$author_n_tracks,"k")  
df_producers1$author_n_tracks = str_remove_all(df_producers1$author_n_tracks,"M") %>%  
  as.numeric()
df_producers1$author_n_tracks = ifelse(df_producers1$author_n_tracks_k==T,df_producers1$author_n_tracks*1000,
                              ifelse(df_producers1$author_n_tracks_m==T, df_producers1$author_n_tracks*1000000,df_producers1$author_n_tracks))
df_producers1 = dplyr::select(df_producers1,-author_n_tracks_k,-author_n_tracks_m)

# Joining datasets into the final one
df_beatstars_top = merge(df_beats1, df_producers1, by="author_name")

# Delete all useless elements from Global Env
rm(df_beats, df_beats1, df_producers, df_producers1)

Data Preparation

# As there are a lot of NAs in beat_key and this variable is not considered to be a good predictor - remove beat_key
# and also beat_title and author_name since there are useless now
df_beatstars_top1 = df_beatstars_top %>% dplyr::select(-beat_key,-author_name,-beat_title)


# Also, a lot of beats does not have MP3 price, so this variable also will be removed
# but it will be done after descriptive statistics part 

# The dependent variable is ranking data with 300+ ranks, so it is hard to choose an appropriate stat model
# Therefore the success coefficient will be made. 
# It falls inside 0 to 1, so the linear regression could be used

df_beatstars_top1$success = ((max(df_beatstars_top1$place) - df_beatstars_top1$place + 1)-0.5) / nrow(df_beatstars_top1)

df_beatstars_top1 = df_beatstars_top1 %>% dplyr::select(-place)

Descriptive Statistics

Beat Naming

Beat is an audio content and not much visual or descriptive can be very used to promote the beat. That is why producers add some features to the title of the beat. These features can offer some beat bundles (Buy N Get N Free) or highlight the low price or just describe the style of the bear (Kid Laroi Type Beat). I am interesting in how really this changes the selling dynamics, so how much beats in top chart have these features.

Highlighting the Price

ggplot(df_beatstars_top1, aes(x=beat_title_price, fill = beat_title_price)) +
  geom_bar(show.legend = FALSE) +
  labs(title = "How many beats has price tag in the title?", x="Price in the Title")

Most of the producers do not use this feature - not enough observations where beat_title_price == TRUE to test if beat_title_price is related to success of the beat.

Special Offer

p1 = ggplot(df_beatstars_top1, aes(x=beat_title_offer, fill = beat_title_offer)) +
  geom_bar(show.legend = FALSE) +
  labs(title = "How many beats has\nspecial offer in the title?") +
  theme(
    axis.title.x = element_blank()
  )

p2 = ggplot(df_beatstars_top1, aes(x=beat_title_offer, y = success, fill = beat_title_offer)) +
  geom_boxplot(show.legend = FALSE) +
  labs(y = "Success Points", title = "\nHow successful are the beats?") +
  theme(
    axis.title.x = element_blank()
  )

grid.arrange(p1, p2,
             widths = c(1, 2),
             bottom = "Offer in the Title")

Some of the producers use this feature, but there are less than a quarter of beats that has special offer in the title.

Judging by the boxplot, offer in the title is not related to the success.

Type Beat

p1 = ggplot(df_beatstars_top1, aes(x = beat_title_type, fill = beat_title_type)) +
  geom_bar(show.legend = FALSE) +
  labs(title = "How many beats has\ntype in the title?") +
  theme(
    axis.title.x = element_blank()
  )

p2 = ggplot(df_beatstars_top1, aes(x=beat_title_type, y = success, fill = beat_title_type)) +
  geom_boxplot(show.legend = FALSE) +
  labs(y = "Success Points", title = "\nHow successful are the beats?") +
  theme(
    axis.title.x = element_blank()
  )

grid.arrange(p1, p2,
             widths = c(1, 2),
             bottom = "Type in the Title")

It seems that the most used feature is to name the type of beat in the title.

Judging by the boxplot, type in the title could be related to the success, so testing it with

shapiro.test(df_beatstars_top1$success) # the DV distribution is not normal

## 
##  Shapiro-Wilk normality test
## 
## data:  df_beatstars_top1$success
## W = 0.95468, p-value = 0.000000002752

leveneTest(df_beatstars_top1$success ~ df_beatstars_top1$beat_title_type) # p>0.05 -> variances are homogeneous

## Levene's Test for Homogeneity of Variance (center = median)
##        Df F value Pr(>F)
## group   1  0.9477 0.3309
##       370

t.test(df_beatstars_top1$success ~ df_beatstars_top1$beat_title_type, var.equal = T) 

## 
##  Two Sample t-test
## 
## data:  df_beatstars_top1$success by df_beatstars_top1$beat_title_type
## t = -1.7849, df = 370, p-value = 0.0751
## alternative hypothesis: true difference in means between group FALSE and group TRUE is not equal to 0
## 95 percent confidence interval:
##  -0.127677637  0.006178105
## sample estimates:
## mean in group FALSE  mean in group TRUE 
##           0.4841593           0.5449091

P-value is a little bit more than 0.05, so there is no sufficient evidence that there is a difference between two groups in terms of success.

Beat Date of Publishing

How many old and new beats are in the chart? Can the beat be relevant in several years from publishing?

ggplot(df_beatstars_top1, aes(x=beat_date)) +
  geom_histogram(bins = 70) +
  labs(x="Date of Publishing")

cat("Correlation coef:", cor(as.numeric(df_beatstars_top1$beat_date),df_beatstars_top1$success))

## Correlation coef: -0.01559687

As expected, most of the beats were published recently, however beats from last year or even from 2020 still can be seen in the top chart. In terms of relationship between success and date of publishing - there is none.

Beat Stats

How many likes, reposts and plays the beat in the top chart usually have?

Likes

describe(df_beatstars_top1$beat_like)

##    vars   n    mean      sd median trimmed    mad min   max range skew kurtosis
## X1    1 372 2114.43 3581.17  737.5 1249.76 913.28   1 29000 28999 3.14     12.6
##        se
## X1 185.67

density <- density(df_beatstars_top1$beat_like)

fig <- plot_ly(x = ~density$x, y = ~density$y, type = 'scatter', mode = 'lines', fill = 'tozeroy')
fig <- fig %>% layout(xaxis = list(title = 'Likes'),
         yaxis = list(title = 'Density'))

fig

cat("Correlation coef:", cor(df_beatstars_top1$beat_like,df_beatstars_top1$success))

## Correlation coef: 0.3543874

Mean=2114, however median = 737, and as it can be seen on graph there are very small amount of beats that has more than 5k likes.

Success and number of likes on the beat are moderately correlated.

Reposts

describe(df_beatstars_top1$beat_repost)

##    vars   n  mean    sd median trimmed   mad min max range skew kurtosis   se
## X1    1 372 35.15 64.07     10   18.98 13.34   0 439   439 3.08    10.81 3.32

density <- density(df_beatstars_top1$beat_repost)

fig <- plot_ly(x = ~density$x, y = ~density$y, type = 'scatter', mode = 'lines', fill = 'tozeroy')
fig <- fig %>% layout(xaxis = list(title = 'Reposts'),
         yaxis = list(title = 'Density'))

fig

cat("Correlation coef:", cor(df_beatstars_top1$beat_repost,df_beatstars_top1$success))

## Correlation coef: 0.3013503

Mean = 35, Median = 10, the dynamics is really similar to Likes

Success and number of reposts on the beat are moderately correlated, it is only a little less than in case of likes.

Plays

describe(df_beatstars_top1$beat_plays)

##    vars   n     mean       sd median  trimmed     mad min     max   range skew
## X1    1 372 112049.8 198849.6  31400 62873.49 40771.5 152 1300000 1299848    3
##    kurtosis       se
## X1     10.2 10309.87

density <- density(df_beatstars_top1$beat_plays)

fig <- plot_ly(x = ~density$x, y = ~density$y, type = 'scatter', mode = 'lines', fill = 'tozeroy')
fig <- fig %>% layout(xaxis = list(title = 'Plays'),
         yaxis = list(title = 'Density'))

fig

cat("Correlation coef:", cor(df_beatstars_top1$beat_plays,df_beatstars_top1$success))

## Correlation coef: 0.3458173

Mean = 112 thousands, median = 31400, the distribution is close to the likes and reposts

Success and number of plays on the beat are moderately correlated, as well as likes and success.

Prices

Common practice is to have different type of licenses for beat purchasing. For one price artist can only use it in non-commercial purposes, for another can have full rights on using beats. Licensing method is always changing in some way, but probably the most common two formats of a beat are MP3(non-tagged) and WAV. So, I am interesting in analyzing the prices among top chart of Beatstars on this two positions

describe(na.omit(df_beatstars_top1$MP3))

##    vars   n  mean   sd median trimmed  mad min max range skew kurtosis   se
## X1    1 311 29.93 7.81  29.99   29.63 1.47   1  55    54 0.38     1.68 0.44

describe(na.omit(df_beatstars_top1$WAV))

##    vars   n  mean    sd median trimmed  mad min    max  range skew kurtosis  se
## X1    1 360 46.27 13.28  49.95   46.08 7.34   3 149.99 146.99 1.33    11.13 0.7

fig1 <- plot_ly(alpha = 0.8)
fig1 <- fig1 %>% add_trace(x = df_beatstars_top1$MP3,
  type='histogram',
  name="MP3 Prices(US $)"
) 
fig1 <- fig1 %>% add_trace(x = df_beatstars_top1$WAV,
  type='histogram',
  name="WAV Prices(US $)"
) 
fig1 <- fig1 %>%
  layout(
    xaxis = list(
      dtick = 10, 
      tick0 = 0, 
      tickmode = "linear"
  ))
fig1 <- fig1 %>% layout(barmode = "overlay")
fig1

Mean Price for MP3 = 29.93 USD, while median = 29.99 USD. Mean Price for WAV = 46.27 USD, while median = 49.95 USD.

So, approximately, WAV version is more expensive than MP3 by 15 to 20 dollars. Also, it can be seen that producers just copy the most appropriate price for the licences, so it is a common thing to have mp3 price = 30 USD and WAV price=50 USD.

Producer Stats

How many plays, followers and tracks the producer in the top chart usually have?

Plays

describe(df_beatstars_top1$author_plays)

##    vars   n    mean      sd  median trimmed     mad min      max    range skew
## X1    1 372 2938817 4464099 1000000 1945698 1339455 282 24300000 24299718 2.91
##    kurtosis       se
## X1     9.56 231452.7

density <- density(df_beatstars_top1$author_plays)

fig <- plot_ly(x = ~density$x, y = ~density$y, type = 'scatter', mode = 'lines', fill = 'tozeroy')
fig <- fig %>% layout(xaxis = list(title = 'Plays'),
         yaxis = list(title = 'Density'))

fig

cat("Correlation coef:", cor(df_beatstars_top1$author_plays,df_beatstars_top1$success))

## Correlation coef: 0.003926138

Mean=2.9M, median =1M. Success and number of plays of the author are not correlated.

Followers

describe(df_beatstars_top1$author_followers)

##    vars   n     mean       sd median  trimmed      mad min    max  range skew
## X1    1 372 49705.81 77571.97  20700 30904.14 26760.93  18 367500 367482 2.54
##    kurtosis      se
## X1     6.35 4021.92

density <- density(df_beatstars_top1$author_followers)

fig <- plot_ly(x = ~density$x, y = ~density$y, type = 'scatter', mode = 'lines', fill = 'tozeroy')
fig <- fig %>% layout(xaxis = list(title = 'Followers'),
         yaxis = list(title = 'Density'))

fig

cat("Correlation coef:", cor(df_beatstars_top1$author_followers,df_beatstars_top1$success))

## Correlation coef: -0.01129017

Mean = 49705, Median = 20700, the dynamics is really similar to Likes. Success and number of followers of the author are not correlated.

Number of Available Beats

describe(df_beatstars_top1$author_n_tracks)

##    vars   n   mean     sd median trimmed   mad min  max range skew kurtosis
## X1    1 372 430.26 425.68    322  360.82 251.3   4 3000  2996 3.29    15.88
##       se
## X1 22.07

density <- density(df_beatstars_top1$author_n_tracks)

fig <- plot_ly(x = ~density$x, y = ~density$y, type = 'scatter', mode = 'lines', fill = 'tozeroy')
fig <- fig %>% layout(xaxis = list(title = 'Number of Available Beats'),
         yaxis = list(title = 'Density'))

fig

cat("Correlation coef:", cor(df_beatstars_top1$author_n_tracks,df_beatstars_top1$success))

## Correlation coef: -0.004064413

The collection of available beats should be huge for top sellers, so it is about experience. Success and number of tracks of the author are not correlated.

Regression Analysis

The dependent variable is “Success” coef. It is continuous and falls inside 0 to 1, so linear regression model could be used.

# remove MP3 Price variable - a lot of NAs
df_beatstars_top2 = df_beatstars_top1 %>% dplyr::select(-MP3) %>% na.omit()

df_beatstars_top2$beat_date = as.numeric(df_beatstars_top2$beat_date)

Correlations between the IVs

library(sjPlot)
tab_corr(df_beatstars_top2[,-c(6:8)], corr.method = "spearman")

	beat_like	beat_repost	beat_date	beat_bpm	beat_plays	WAV	author_followers	author_plays	author_n_tracks	success
beat_like		0.933***	-0.737***	0.081	0.993***	-0.069	0.292***	0.299***	0.093	0.404***
beat_repost	0.933***		-0.648***	0.067	0.926***	-0.049	0.253***	0.261***	0.050	0.389***
beat_date	-0.737***	-0.648***		-0.001	-0.750***	0.042	-0.116*	-0.090	-0.044	-0.096
beat_bpm	0.081	0.067	-0.001		0.079	-0.113*	0.039	-0.015	-0.001	0.136**
beat_plays	0.993***	0.926***	-0.750***	0.079		-0.089	0.307***	0.319***	0.102	0.388***
WAV	-0.069	-0.049	0.042	-0.113*	-0.089		-0.112*	-0.097	-0.008	-0.093
author_followers	0.292***	0.253***	-0.116*	0.039	0.307***	-0.112*		0.918***	0.598***	0.017
author_plays	0.299***	0.261***	-0.090	-0.015	0.319***	-0.097	0.918***		0.641***	0.032
author_n_tracks	0.093	0.050	-0.044	-0.001	0.102	-0.008	0.598***	0.641***		-0.073
success	0.404***	0.389***	-0.096	0.136**	0.388***	-0.093	0.017	0.032	-0.073
Computed correlation used spearman-method with listwise-deletion.

The success is only correlated with beat_like, beat_repost, beat_plays and beat_bpm, however beat_like, beat_repost, beat_plays are highly correlated variables, so I will choose only beat_like and beat_bpm as a predictors in the model.

Model Assumptions (diagnostics)

# model 1
m1 = lm(success ~ beat_like + beat_bpm, data = df_beatstars_top1)

# check assumptions m1
layout(matrix(c(1,2,3,4),2,2)) 
plot(m1)

cat("Assumptions of model 1 are violated, but it is not so bad")

## Assumptions of model 1 are violated, but it is not so bad

# outliers
outlierTest(m1)

## No Studentized residuals with Bonferroni p < 0.05
## Largest |rstudent|:
##      rstudent unadjusted p-value Bonferroni p
## 263 -2.457281            0.01446           NA

cat("Model 1: no outliers")

## Model 1: no outliers

# influential points
cutoff <- 4/((nrow(df_beatstars_top1)-length(m1$coefficients)-2)) 
plot(m1, which=4, cook.levels=cutoff) 
cat("Model 1: 180,263,17 are influential points")

## Model 1: 180,263,17 are influential points

# Homoscedasticity Assumption
ncvTest(m1) # assumption is not violated

## Non-constant Variance Score Test 
## Variance formula: ~ fitted.values 
## Chisquare = 0.5019704, Df = 1, p = 0.47864

cat("Model 1: Homoscedasticity Assumption is not violated")

## Model 1: Homoscedasticity Assumption is not violated

# fix ds
df_beatstars_top2 = df_beatstars_top1[-c(180,263,17),]

# model 2
m2 = lm(success ~ beat_like + beat_bpm, data = df_beatstars_top2)

cat("Fixing model 1, the assumptions plot of fixed model")

## Fixing model 1, the assumptions plot of fixed model

layout(matrix(c(1,2,3,4),2,2)) 

plot(m2)

cat("Quite okay right now")

## Quite okay right now

Model interpretation

summary(m2)

## 
## Call:
## lm(formula = success ~ beat_like + beat_bpm, data = df_beatstars_top2)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -0.60579 -0.21417  0.00606  0.20985  0.48943 
## 
## Coefficients:
##                Estimate  Std. Error t value            Pr(>|t|)    
## (Intercept) 0.321132740 0.039462867   8.138 0.00000000000000637 ***
## beat_like   0.000034514 0.000004339   7.954 0.00000000000002271 ***
## beat_bpm    0.000984919 0.000318694   3.090             0.00215 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.2645 on 366 degrees of freedom
## Multiple R-squared:  0.1636, Adjusted R-squared:  0.1591 
## F-statistic: 35.81 on 2 and 366 DF,  p-value: 0.000000000000006276

16% of variance was explained by the model, which says that fit is okay for creative market and only 2 predictors.

Both predictors are significant and have a positive relationship with success, however the effect is incredibly small for both of the predictors.

So, definitely, the more likes (or reposts or plays) on the beat and the more bpm of the beat the higher is the success of the beat, but the effect is barely noticeable.