1. Load and clean the data

Your three weekly files are read together. Things this section deals with (based on how the files are actually structured):

  • Duplicates – the weekly files overlap almost completely, so the same post (uri) appears in several files. Each post is kept once (the newest file’s engagement counts win).
  • Missing created_at – some rows have NA; the timestamp is recovered from the post_data column when possible.
  • Empty posts – image/quote-only posts with no text are dropped from text analysis.
  • Mixed timestamp precision – handled by lubridate::ymd_hms().
files <- list.files(params$data_dir, pattern = "\\.csv$", recursive = TRUE, full.names = TRUE)
stopifnot("No CSV files found - check params$data_dir" = length(files) > 0)

raw <- map_dfr(files, function(f) {
  read_csv(f, col_types = cols(.default = col_character()), show_col_types = FALSE) %>%
    mutate(source_file = basename(f))
})

num_cols <- c("reply_count", "repost_count", "like_count", "quote_count", "bookmark_count")

posts_all <- raw %>%
  mutate(
    # fall back to the timestamp stored inside post_data when created_at is NA
    created_chr = coalesce(
      created_at,
      str_extract(post_data, "\\d{4}-\\d{2}-\\d{2}T\\d{2}:\\d{2}:\\d{2}(\\.\\d+)?")
    ),
    created = ymd_hms(created_chr, tz = "UTC", quiet = TRUE),
    across(all_of(num_cols), ~ replace_na(suppressWarnings(as.numeric(.x)), 0))
  ) %>%
  arrange(desc(source_file)) %>%          # newest file first ...
  distinct(uri, .keep_all = TRUE)         # ... so its counts are the ones kept

clean_text <- function(x) {
  x %>%
    str_replace_all("[‘’]", "'") %>%
    str_replace_all("https?://\\S+|www\\.\\S+", " ") %>%                      # full URLs
    str_replace_all("\\b[\\w-]+(\\.[\\w-]+)*\\.(com|gg|net|org|io|co|tv|uk|de|social)\\S*", " ") %>% # truncated links
    str_replace_all("@[\\w.-]+", " ") %>%                                     # @mentions
    str_replace_all("#", " ") %>%                                             # keep hashtag word, drop '#'
    str_to_lower() %>%
    str_squish()
}

posts <- posts_all %>%
  filter(!is.na(text), str_squish(text) != "") %>%
  { if (params$english_only) filter(., is.na(langs) | str_detect(langs, "^en")) else . } %>%
  filter(!is.na(created)) %>%
  mutate(
    date       = as_date(created),
    period     = as_date(floor_date(created, params$time_unit, week_start = 1)),
    text_clean = clean_text(text),
    engagement = like_count + repost_count + reply_count + quote_count
  )

tibble(
  Step  = c("Rows read (all files)", "Unique posts", "Posts used in analysis"),
  Count = c(nrow(raw), nrow(posts_all), nrow(posts))
) %>% kable(caption = "Data cleaning summary")
Data cleaning summary
Step Count
Rows read (all files) 573
Unique posts 192
Posts used in analysis 169

2. Overview

tibble(
  Metric = c("Posts", "Unique authors", "First post", "Last post",
             "Total likes", "Median likes per post"),
  Value  = c(nrow(posts), n_distinct(posts$author_handle),
             as.character(min(posts$date)), as.character(max(posts$date)),
             comma(sum(posts$like_count)), median(posts$like_count))
) %>% kable()
Metric Value
Posts 169
Unique authors 136
First post 2026-01-29
Last post 2026-11-10
Total likes 781
Median likes per post 1

Posting volume over time

volume <- posts %>%
  count(period, name = "n") %>%
  complete(period = seq(min(period), max(period), by = params$time_unit),
           fill = list(n = 0))

plot_ly(volume, x = ~period, y = ~n, type = "bar", marker = list(color = pal$main),
        hovertemplate = "%{x|%b %d, %Y}<br>%{y} posts<extra></extra>") %>%
  layout(title = list(text = paste("Posts per", params$time_unit), x = 0),
         xaxis = list(title = ""), yaxis = list(title = "Posts"))

When do people post?

day_names <- c("Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun")

rhythm <- posts %>%
  mutate(local = with_tz(created, params$tz),
         weekday = day_names[wday(local, week_start = 1)],
         hour = hour(local)) %>%
  count(weekday, hour)

plot_ly(rhythm, x = ~hour, y = ~weekday, z = ~n, type = "heatmap", colors = "Blues",
        hovertemplate = "%{y} %{x}:00<br>%{z} posts<extra></extra>") %>%
  layout(title = list(text = paste0("Posting rhythm (", params$tz, ")"), x = 0),
         xaxis = list(title = "Hour of day", dtick = 2, range = c(-0.5, 23.5)),
         yaxis = list(title = "", categoryorder = "array",
                      categoryarray = day_names, autorange = "reversed"))

Most active accounts

posts %>%
  count(author_handle, sort = TRUE) %>%
  slice_head(n = 15) %>%
  hbar("author_handle", "n", "Most active accounts", xlab = "Posts")

3. Basic text analysis

# Edit this list freely: words that are in every post carry no information.
custom_stop <- c("call", "duty", "cod", "game", "games", "gaming",
                 "amp", "http", "https", "www", "com", "bsky", "social", "rt")

tokens <- posts %>%
  select(uri, author_handle, created, period, text_clean) %>%
  unnest_tokens(word, text_clean) %>%
  mutate(word = str_remove(word, "'s$")) %>%
  filter(str_detect(word, "^[a-z][a-z']+$"),
         nchar(word) > 2,
         !word %in% stop_words$word,
         !word %in% custom_stop)

word_counts <- count(tokens, word, sort = TRUE)

Most frequent words

word_counts %>%
  slice_head(n = params$top_n) %>%
  hbar("word", "n", paste("Top", params$top_n, "words"))

Most frequent word pairs

junk <- c(stop_words$word, "http", "https", "www", "com", "amp")

bigrams <- posts %>%
  select(uri, text_clean) %>%
  unnest_tokens(bigram, text_clean, token = "ngrams", n = 2) %>%
  filter(!is.na(bigram)) %>%
  separate(bigram, c("w1", "w2"), sep = " ") %>%
  filter(str_detect(w1, "^[a-z]+$"), str_detect(w2, "^[a-z]+$"),
         !w1 %in% junk, !w2 %in% junk) %>%
  unite(bigram, w1, w2, sep = " ") %>%
  distinct(uri, bigram)                       # count each bigram once per post

bigrams %>%
  count(bigram, sort = TRUE) %>%
  slice_head(n = params$top_n) %>%
  hbar("bigram", "n", "Most frequent word pairs", color = "#7b5ea7", xlab = "Posts")

Hashtags

hashtags <- posts %>%
  mutate(tags = replace_na(tags, "")) %>%
  transmute(uri, period,
            tag = map2(text, tags, function(txt, tg) {
              c(str_remove(str_extract_all(txt, "#\\w+")[[1]], "#"),
                str_split(tg, ";\\s*")[[1]])
            })) %>%
  unnest(tag) %>%
  mutate(tag = str_to_lower(str_trim(tag))) %>%
  filter(tag != "") %>%
  distinct(uri, tag, .keep_all = TRUE)

if (nrow(hashtags) > 0) {
  hashtags %>% count(tag, sort = TRUE) %>% slice_head(n = params$top_n) %>%
    hbar("tag", "n", "Most used hashtags", color = "#c2549d", xlab = "Posts")
} else {
  cat("No hashtags found.")
}

5. Sentiment analysis

Sentiment uses the Bing lexicon (each word is tagged positive or negative; no download needed). A post’s net score = positive words − negative words. Posts with no lexicon words score 0 and count as “Neutral”.

Domain caveat: lexicons are built for everyday English. In gaming talk, words like war, strike, shot, kill are not emotional, so a short list is removed below. Edit sentiment_exclude to suit your topic.

sentiment_exclude <- c("like", "well", "war", "warfare", "strike", "shot",
                       "ghost", "kill", "killed", "killing", "gun", "bomb")

bing <- get_sentiments("bing") %>%
  filter(!word %in% sentiment_exclude) %>%
  distinct(word, .keep_all = TRUE)

sent_by_post <- tokens %>%
  inner_join(bing, by = "word") %>%
  count(uri, sentiment) %>%
  complete(uri, sentiment = c("positive", "negative"), fill = list(n = 0)) %>%
  pivot_wider(names_from = sentiment, values_from = n, values_fill = 0)

posts_sent <- posts %>%
  left_join(sent_by_post, by = "uri") %>%
  mutate(across(c(positive, negative), ~ replace_na(.x, 0)),
         net   = positive - negative,
         label = case_when(net > 0 ~ "Positive", net < 0 ~ "Negative", TRUE ~ "Neutral"),
         label = factor(label, levels = c("Positive", "Neutral", "Negative")))

posts_sent %>% count(label) %>% mutate(share = percent(n / sum(n), 0.1)) %>% kable()
label n share
Positive 48 28.4%
Neutral 62 36.7%
Negative 59 34.9%

Overall distribution

dist <- posts_sent %>% count(net)

plot_ly(dist, x = ~net, y = ~n, type = "bar",
        marker = list(color = ifelse(dist$net > 0, pal$pos, ifelse(dist$net < 0, pal$neg, pal$neu))),
        hovertemplate = "Net %{x}: %{y} posts<extra></extra>") %>%
  layout(title = list(text = "Distribution of post sentiment (net Bing score)", x = 0),
         xaxis = list(title = "Net score (negative ← 0 → positive)", dtick = 1),
         yaxis = list(title = "Posts"))

Words driving the sentiment

sent_words <- tokens %>%
  inner_join(bing, by = "word") %>%
  count(word, sentiment) %>%
  group_by(sentiment) %>%
  slice_max(n, n = 10, with_ties = FALSE) %>%
  ungroup() %>%
  mutate(signed = ifelse(sentiment == "negative", -n, n))

plot_ly(sent_words, x = ~signed, y = ~word, color = ~sentiment, text = ~n,
        colors = c(negative = pal$neg, positive = pal$pos),
        type = "bar", orientation = "h",
        hovertemplate = "%{y}: %{text} mentions<extra></extra>") %>%
  layout(title = list(text = "Top positive and negative words", x = 0),
         xaxis = list(title = "← negative  |  positive →"),
         yaxis = list(title = "", categoryorder = "array",
                      categoryarray = sent_words$word[order(sent_words$signed)]),
         margin = list(l = 110))

Sentiment over time

sent_trend <- posts_sent %>%
  group_by(period) %>%
  summarise(n = n(),
            mean_net = mean(net),
            se = sd(net) / sqrt(n()),
            Positive = mean(label == "Positive"),
            Neutral  = mean(label == "Neutral"),
            Negative = mean(label == "Negative"),
            .groups = "drop") %>%
  filter(n >= params$min_n)

# Average net sentiment (error bars = +/- 1 standard error)
plot_ly(sent_trend, x = ~period, y = ~mean_net, type = "scatter", mode = "lines+markers",
        line = list(color = pal$main), marker = list(color = pal$main),
        error_y = list(array = ~se, color = "#9bb3d6"),
        text = ~paste(n, "posts"),
        hovertemplate = "%{x|%b %d, %Y}<br>Mean net: %{y:.2f}<br>%{text}<extra></extra>") %>%
  layout(title = list(text = paste("Average sentiment per", params$time_unit), x = 0),
         xaxis = list(title = ""),
         yaxis = list(title = "Mean net score", zeroline = TRUE),
         shapes = list(list(type = "line", x0 = 0, x1 = 1, xref = "paper",
                            y0 = 0, y1 = 0, line = list(color = "#bbb", dash = "dot"))))
# Share of positive / neutral / negative posts
plot_ly(sent_trend, x = ~period, y = ~Positive, type = "bar", name = "Positive",
        marker = list(color = pal$pos)) %>%
  add_trace(y = ~Neutral,  name = "Neutral",  marker = list(color = pal$neu)) %>%
  add_trace(y = ~Negative, name = "Negative", marker = list(color = pal$neg)) %>%
  layout(barmode = "stack",
         title = list(text = "Sentiment mix over time", x = 0),
         xaxis = list(title = ""),
         yaxis = list(title = "Share of posts", tickformat = ".0%"))

Most positive and most negative posts

show_cols <- function(df) {
  df %>% transmute(Date = date, Account = author_handle, Net = net,
                   Post = str_trunc(str_squish(text), 180))
}

posts_sent %>% slice_max(net, n = 5, with_ties = FALSE) %>% show_cols() %>%
  kable(caption = "Most positive")
Most positive
Date Account Net Post
2026-07-05 landht.bsky.social 5 Hogwarts Legacy 2 Perfect Setting And Era Has Fans More Excited Than Ever Hogwarts Legacy came out three years ago, and the ripple effects of its success are still being felt. I…
2026-10-23 byteseu.bsky.social 4 Call of Duty: Modern Warfare 4 tops Steam Korea sales at launch https://www.byteseu.com/2473905/ Call of Duty: Modern Warfare 4 debuted at No. 1 in Korea as its North Korean inv…
2026-10-16 shawntheair.bsky.social 3 Hot Take: Sales wise, this is probably true. But these games are also multiplatform and have larger install bases. However, last year’s Call of Duty was panned by fans and criti…
2026-06-15 brianpshea.bsky.social 3 The fact that a studio can win several awards just last year and still get shut down because it doesn’t put up Call of Duty sales numbers says a lot about this industry’s curren…
2026-04-23 thesocialsage.bsky.social 3 Analysts say Xbox Game Pass price cut is ‘not surprising at all’ since Call of Duty’s addition to the service ’did not lead to a significant increase in Xbox console sales or ev…
posts_sent %>% slice_min(net, n = 5, with_ties = FALSE) %>% show_cols() %>%
  kable(caption = "Most negative")
Most negative
Date Account Net Post
2026-09-22 jcpxdesigns.bsky.social -4 I remember when it first started I thought, oh finally someone is just as annoyed as me at the gaming press ignoring abuses in the game industry like excessive crunch, constant …
2026-09-20 gamespur.games -4 Imagine paying actual money to ruin everyone’s lobby. The state of this industry is truly chaotic, but this breakdown is honestly wild. https://www.eurogamer.net/activision-vide…
2026-09-20 bestdealssnews.bsky.social -4 Call of Duty maker Activision claims video game cheating is now a £6.3 billion ($8.5 billion) a year industry Activision has published a breakdown of the numbers behind cheating…
2026-03-29 gallingcrow.bsky.social -4 i have no idea how Arc raiders became such a success. it feels like shit and looks indistinguishable from a call of duty directed by Neil Blomkamp just a viciously ugly game, i …
2026-03-24 gigowat.bsky.social -4 Windows Central より Call of Duty: Black Ops 7 is still selling ridiculously well, despite being a ‘dead’ game everyone seems to hate コール・オブ・デューティ:ブラック・オプス7は、誰もが憎むように見える「死んだ」ゲームにも…
# Optional second opinion: AFINN scores words from -5 to +5 (captures intensity).
# First run only: install.packages("textdata"); then accept the one-time download prompt.
afinn <- get_sentiments("afinn") %>% filter(!word %in% sentiment_exclude)

afinn_post <- tokens %>%
  inner_join(afinn, by = "word") %>%
  group_by(uri) %>%
  summarise(afinn = sum(value), .groups = "drop")

afinn_trend <- posts %>%
  left_join(afinn_post, by = "uri") %>%
  mutate(afinn = replace_na(afinn, 0)) %>%
  group_by(period) %>%
  summarise(n = n(), mean_afinn = mean(afinn), .groups = "drop") %>%
  filter(n >= params$min_n)

plot_ly(afinn_trend, x = ~period, y = ~mean_afinn, type = "scatter", mode = "lines+markers",
        line = list(color = "#7b5ea7"), text = ~paste(n, "posts"),
        hovertemplate = "%{x|%b %d, %Y}<br>Mean AFINN: %{y:.2f}<br>%{text}<extra></extra>") %>%
  layout(title = list(text = "AFINN sentiment over time", x = 0),
         xaxis = list(title = ""), yaxis = list(title = "Mean AFINN score"))

6. Engagement and sentiment

Engagement over time

eng_trend <- posts %>%
  group_by(period) %>%
  summarise(n = n(), mean_likes = mean(like_count), median_likes = median(like_count),
            .groups = "drop") %>%
  filter(n >= params$min_n)

plot_ly(eng_trend, x = ~period, y = ~mean_likes, type = "scatter", mode = "lines+markers",
        name = "Mean likes", line = list(color = pal$main), text = ~paste(n, "posts"),
        hovertemplate = "%{x|%b %d, %Y}<br>%{y:.1f} likes (%{text})<extra></extra>") %>%
  add_trace(y = ~median_likes, name = "Median likes",
            line = list(color = pal$neu, dash = "dash")) %>%
  layout(title = list(text = paste("Likes per post, by", params$time_unit), x = 0),
         xaxis = list(title = ""), yaxis = list(title = "Likes per post"))

Do positive or negative posts get more engagement?

plot_ly(posts_sent, x = ~label, y = ~log1p(engagement), color = ~label,
        colors = c(Positive = pal$pos, Neutral = pal$neu, Negative = pal$neg),
        type = "box", boxpoints = "all", jitter = 0.4, pointpos = 0,
        text = ~str_trunc(str_squish(text), 80),
        hovertemplate = "%{text}<br>Engagement: %{customdata}<extra></extra>",
        customdata = ~engagement) %>%
  layout(title = list(text = "Engagement by sentiment (log scale)", x = 0),
         xaxis = list(title = ""), yaxis = list(title = "log(1 + likes + reposts + replies + quotes)"),
         showlegend = FALSE)
ct <- cor.test(posts_sent$net, posts_sent$engagement, method = "spearman", exact = FALSE)

Spearman correlation between net sentiment and engagement: rho = -0.156, p = 0.0426.

Most engaging posts

posts_sent %>%
  slice_max(engagement, n = 10, with_ties = FALSE) %>%
  transmute(Date = date, Account = author_handle, Engagement = engagement,
            Sentiment = as.character(label), Post = str_trunc(str_squish(text), 160)) %>%
  kable()
Date Account Engagement Sentiment Post
2026-02-02 queerlybee.bsky.social 203 Neutral I’ve seen an uptick in people streaming minecraft lately & i just want to remind folks that its one of the games listed on the BDS list to boycott for palest…
2026-06-15 brianpshea.bsky.social 138 Positive The fact that a studio can win several awards just last year and still get shut down because it doesn’t put up Call of Duty sales numbers says a lot about th…
2026-08-07 matpiscatella.bsky.social 83 Positive Yes, Grand Theft Auto V ranks 1st in both full game unit and dollar sales in the US over the past decade, while Red Dead Redemption II ranks 3rd in each. Cal…
2026-09-25 alexaray.bsky.social 49 Negative The year all of this popped off publicly was the year the Call of Duty I wrote on, Call of Duty: Vanguard, came out. That year, the press and gamer communiti…
2026-04-13 irishdocsenv.bsky.social 40 Negative Updated Statement on Government Response to Fossil Fuel Blockades. The government is failing to deliver a Just Transition and instead binding us to further f…
2026-08-14 jakesteinberg.substack.com 34 Neutral “Good news! You can buy South of Midnight now because your money won’t go to Xbox Anyway, here are a bunch of Xbox games and controllers to spend your money …
2026-10-17 gnarlyanimal.bsky.social 32 Negative we literally have so many good video games coming out recently and we’re still seeing little think pieces about how gaming is dead written by niggas with 600…
2026-06-15 andyrobinson.bsky.social 29 Negative Last year, Compulsion CEO Guillaume Provost said the studio’s role within Microsoft was“not to make blockbusters like Call of Duty, but to provide games that…
2026-02-27 stephentotilo.bsky.social 28 Negative NEW: Major League Baseball took issue this week with the name of Call of Duty’s “World Series of Warzone” esports series, citing potential brand confusion Ac…
2026-10-15 deanpreston.bsky.social 25 Negative This Y Combinator backed tech bro is inciting violence against people in the Tenderloin. This is one of the most disgusting things I’ve seen in San Francisco…

Notes and limitations

  • Small sample. With a few hundred posts, monthly averages move a lot; the error bars and the min_n filter are there to keep you honest. Check the post counts before reading anything into a trend.
  • Lexicon sentiment is a blunt tool. It misses sarcasm, negation (“not good”) and gaming slang. Treat results as directional; for stronger claims, hand-code a sample of posts and compare.
  • Search bias. These posts come from your scrape query, so trends reflect what the query captured, not all of Bluesky.
  • Timestamps use created_at (UTC), falling back to the timestamp in post_data.
  • To change the bucket size, knit with e.g. rmarkdown::render("bluesky_analysis.Rmd", params = list(time_unit = "week")).