# pyscience imports
import numpy as np
import pandas as pd
pd.set_option('float_format', '{:f}'.format)

# viz
import matplotlib.pyplot as plt
import plotnine as pn

font = {'family' : 'IBM Plex Sans',
               'weight' : 'normal',
               'size'   : 10}
plt.rc('font', **font)

_id : unique identifier combining the sale’s URL and UTC timestamp.
url : the path to the item on Bandcamp. Use this column to join this dataset to the dataset of Bandcamp items.
artist_name : Name of the artist.
album_title : Title of the album, if applicable.
art_url : path to the item’s art image.
item_type : denotes the type of object. a for digital albums, p for physical items, and t for digital tracks.
slug_type : also denotes the type of object. a for all albums, p for merch, and t for tracks.
utc_date : the UTC timestamp of the sale datetime.
country_code : country code of the buyer.
country : full country code name of the buyer.
item_price : price of the item in the seller’s currency.
currency : the seller’s currency.
amount_paid : amount paid in the seller’s currency.
amount_paid_fmt : amount paid in the seller’s currency, with the currency symbol.
amount_paid_usd : amount paid converted to US Dollars.
amount_over_fmt : amount voluntarily paid over the item price in the seller’s currency.

dataroot = "/home/alal/tmp/scratch_data/000_musicData/"
df = pd.read_csv(dataroot+"1000000-bandcamp-sales.csv").infer_objects()

df['utc_date2'] = pd.to_datetime(df['utc_date'], unit='s')
df['date'], df['hour'] = df.utc_date2.dt.date, df.utc_date2.dt.strftime('%H')
df['country2'] = np.where(df.country.isin(df.country.value_counts()[:10].index), df.country, 'Other')
df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1000000 entries, 0 to 999999
Data columns (total 28 columns):
 #   Column                 Non-Null Count    Dtype         
---  ------                 --------------    -----         
 0   Unnamed: 0             1000000 non-null  int64         
 1   _id                    1000000 non-null  object        
 2   art_url                1000000 non-null  object        
 3   item_type              1000000 non-null  object        
 4   utc_date               1000000 non-null  float64       
 5   country_code           1000000 non-null  object        
 6   track_album_slug_text  2237 non-null     object        
 7   country                1000000 non-null  object        
 8   slug_type              988416 non-null   object        
 9   amount_paid_fmt        1000000 non-null  object        
 10  item_price             1000000 non-null  float64       
 11  item_description       999977 non-null   object        
 12  art_id                 764520 non-null   float64       
 13  url                    1000000 non-null  object        
 14  amount_paid            1000000 non-null  float64       
 15  releases               11584 non-null    float64       
 16  artist_name            999990 non-null   object        
 17  currency               1000000 non-null  object        
 18  album_title            356461 non-null   object        
 19  amount_paid_usd        1000000 non-null  float64       
 20  package_image_id       232708 non-null   float64       
 21  amount_over_fmt        119133 non-null   object        
 22  item_slug              21742 non-null    object        
 23  addl_count             21742 non-null    float64       
 24  utc_date2              1000000 non-null  datetime64[ns]
 25  date                   1000000 non-null  object        
 26  hour                   1000000 non-null  object        
 27  country2               1000000 non-null  object        
dtypes: datetime64[ns](1), float64(8), int64(1), object(18)
memory usage: 213.6+ MB

EDA

grouped_day_level = df.groupby(['country2', 'date']).size().reset_index(name='count')
(
    pn.ggplot(grouped_day_level, pn.aes(x="date", y="count"))
    + pn.geom_point()
    + pn.facet_wrap("country2", scales="free_y")
    + pn.theme_matplotlib()
    + pn.theme(figure_size=(16, 8), axis_text_x=pn.element_text(angle=90))
)

Strong weekend effects.

grouped_hour_level = df.groupby(['country2', 'hour']).size().reset_index(name='count')
(
    pn.ggplot(grouped_hour_level, pn.aes(x='hour', y="count"))
    + pn.geom_point()
    + pn.facet_wrap("country2", scales="free_y")
    + pn.theme_matplotlib()
    + pn.theme(figure_size=(16, 8),
            axis_text_x=pn.element_text(angle=90)
        )
)

Peaks during the workday (~9 AM).

bandcamp fridays (‘RD in Time’)

We bin the data into 10 minute windows for a high-resolution look around the midnight pacific announcement.

bcf = (df
    .query("utc_date2.dt.date == @pd.to_datetime('2020-10-02').date()")
    .assign(time_k_min=lambda x: x['utc_date2'].dt.floor('10min'))
    .groupby(['country2', 'time_k_min']).size().reset_index(name='count')
)
bcf['time_label'] = bcf['time_k_min'].dt.strftime('%H:%M')
bcf['hours'] = bcf['time_k_min'].dt.hour + bcf['time_k_min'].dt.minute / 60

# intensive margin
df['pct_over'] = np.where(df.item_price > 0, (df.amount_paid - df.item_price) / df.item_price, 1)
df['pct_over'] = np.clip(df['pct_over'], 0, 1) # cap at 100%
gener = (
    df
    .query("utc_date2.dt.date == @pd.to_datetime('2020-10-02').date()")
    .assign(time_k_min=lambda x: x['utc_date2'].dt.floor('10min'))
    .groupby(['country2', 'time_k_min']).pct_over.mean().reset_index(name='generosity')
)
bcf = bcf.merge(gener, on=['country2', 'time_k_min'])
bcf['log1p_generosity'] = np.log1p(bcf['generosity'])

# Create plots
f1 = (
    pn.ggplot(bcf, pn.aes(x="hours", y="count"))
    + pn.geom_point()
    + pn.geom_vline(xintercept=7-.01, linetype="dashed", color="red")
    + pn.facet_wrap("country2", scales="free_y")
    + pn.theme_bw()
    + pn.theme(
        figure_size=(16, 8),
        axis_text_x=pn.element_text(angle=90),
        dpi=200,
    )
    + pn.labs(
        title="Users response to Bandcamp Friday: Evidence from 1M transactions",
        subtitle="Extensive Margin",
        y="sales (count)",
        x="Hour (UTC), Announcements go out at midnight pacific (7 UTC)",
        caption="Transaction level data from https://components.one/datasets/bandcamp-sales",
    )
    + pn.scale_x_continuous(
        breaks=range(0, 25), labels=lambda l: [f"{int(x):02d}:00" for x in l]
))

f1

Very strong effect across all countries, including (perhaps surprisingly), the US and Canada, where the the bump in transactions is likely driven by midnight shoppers on the west coast.

# Create the plot
f2 = (
    pn.ggplot(bcf, pn.aes(x="hours", y="log1p_generosity"))
    + pn.geom_point()
    + pn.geom_vline(xintercept=7-.01, linetype="dashed", color="red")
    + pn.facet_wrap("country2", scales="free_y")
    + pn.theme_bw()
    + pn.theme(
        figure_size=(16, 8),
        axis_text_x=pn.element_text(angle=90),
        dpi=200,
    )
    + pn.labs(
        title="Users response to Bandcamp Friday: Evidence from 1M transactions",
        subtitle="Intensive Margin",
        y="Percent over asking price",
        x="Hour (UTC)",
        caption="Transaction level data from https://components.one/datasets/bandcamp-sales",
    )
    + pn.scale_x_continuous(
        breaks=range(0, 25), labels=lambda l: [f"{int(x):02d}:00" for x in l]
    )
)
f2

Since buyers can set their own price, we can compute their generosity in terms of the percentage over the asking price. Generosity doesn’t change much around the annoucement. So we conclude that the bandcamp friday effect is driven by the extensive margin (transactions) rather than the intensive margin (generosity).