The Most Unexpectedly Good and Bad TV Episodes

Creating an episode level data set

So far the basics_agg data set is just a list of 700 TV Series and their information. To build a model to predict episode ratings I’ll have to build a data set where each row is an episode. This will replicate some the logic from above that merges the 3 data-sets together:

all_tv_details <- basics_agg %>% 
  ## Join in Episode Data
  inner_join(episodes, by = join_by(tconst==parentTconst)) %>%
  ## Join in Ratings Data
  inner_join(ratings, by = join_by(tconst.y == tconst)) %>% 
  # Bring in Episode Titles
  left_join(basics %>% filter(titleType == 'tvEpisode') %>% 
              transmute(tconst, episodeTitle = primaryTitle),
            by = join_by(tconst.y == tconst)) %>% 
  arrange(tconst, seasonNumber, episodeNumber) %>% 
  group_by(tconst) %>% 
  # Create variables for total number episodes
  mutate(episodeOverall = row_number(tconst),
         seasonNumber = factor(seasonNumber)
  ) %>%
  # Filter Out Missing Data
  filter(!is.na(seasonNumber) & !is.na(episodeNumber)) %>%
  ungroup()

Now the dataset is all prepared to find our unexpected episodes.

Function to find the Unexpected Episodes

The steps above have been built into a function called get_anomalies() which runs the four steps described above. The parameter onlyAnomalies determines whether to return only unexpected episodes or to return all episodes. The differences described in step 2 are added using the augment function from broom :

get_anomalies <- function(dt, onlyAnomalies = T){
  
  ## STEP 1: Run Linear Model on IMDB Ratings vs. Episode Number + Season Info
  #if multiple seasons for show use both global and local trend
  if(n_distinct(dt$seasonNumber) > 1){
    model <- lm(averageRating ~ episodeOverall + seasonNumber*episodeNumber, 
                data = dt)
  }
  # if only one season then global trend = local trend
  else{
    model <- lm(averageRating ~ episodeOverall, data = dt)
  }
  
  ### Step 2 - Add in Residuals from model to initial data set
  results <- augment(model, dt) %>% 
  ### Step 3 - Calculate the 3*IQR Range for each episode
    mutate(
      ## Determine the IQR of the Residuals (P75 - P25)
      iqr = (quantile(.resid, .75)-quantile(.resid, .25)),
      ## Set Lower Bound for expected range of residuals
      lci = quantile(.resid, .25)-(3*iqr),
      ## Set Upper Bound for expected range of residuals
      uci = quantile(.resid, .75)+(3*iqr),
      ## Tag an episode as an anomaly if its actual rating is outside the bounds
      anomaly = if_else(.resid > uci | .resid < lci, T, F),
      
      ## Set expected range back in the scale of the 0-10 prediction.
      lower = .fitted + lci,
      upper = .fitted + uci,
      
      # Step 4 - Calculate the difference between the bounds and the actual 
      # value to use for measure of unexpectedness
      remainder = if_else(.resid < 0, averageRating-lower, averageRating-upper)
    ) %>% 
    # Subset columns
    select(episodeOverall, seasonNumber, episodeNumber, episodeTitle, 
           averageRating, .fitted, .resid, 
           anomaly, lower, upper, remainder)
  
  # Determine whether to return all episodes or just the unexpected episodes
  if(onlyAnomalies == T){
    return(results %>% filter(anomaly == T))
  }
  else{
    return(results)
  }
}

The TV Shows with the Most Unexpected Episodes

Overall, 143 episodes were identified as being unexpectedly good or bad. Of this bunch 64% are unexpectedly bad showing that its more common for a good show to miss then it is for a show to hit an unexpected home run.

The first thing I want to look at are the 10 TV Series that have the most unexpected episodes both good and bad.

results %>% 
  # Group by Show
  group_by(primaryTitle) %>% 
  # Count the number of unexpected episodes in total as well as good and bad
  summarize(
    total = n(),
    `unexpectedly bad` = sum(.resid < 0)*-1,
    `unexpectedly good` = sum(.resid > 0)
  ) %>%
  # Get the Top 10 by Total
  slice_max(order_by = total, n = 10, with_ties = F) %>% 
  pivot_longer(
    cols = c(`unexpectedly bad`, `unexpectedly good`),
    names_to = "type",
    values_to = "episodes"
  ) %>% 
  ggplot(aes(x = episodes, y=fct_reorder(primaryTitle, total), fill = type)) + 
    geom_col() + 
    geom_text(aes(label = if_else(episodes != 0, abs(episodes), NA)), 
              hjust = "inward", color = 'grey90') +
    labs(x = "# of Unexpected Episodes",
         y = "TV Series",
         title = "TV Series with Most <i style = 'color:#ba2a22'>Unexpected</i> Episodes",
         fill = "") + 
    scale_fill_viridis_d(option = "cividis", begin = .2, end = .8, 
                         labels = str_to_title) +
     
    theme(
      legend.position = 'top',
      plot.title = element_markdown(),
      axis.text.x = element_blank(),
      axis.ticks.x = element_blank(),
      axis.title = element_text(size = 12),
      axis.text.y = element_text(size = 10),
      legend.margin = margin(0, 0, -5, 0)
    )

Surprisingly, at least to me, SpongeBob Squarepants has the most unexpected episodes with 5 and they’re all unexpectedly bad. This is in contrast with Desperate Housewives and Big Bang Theory which have all unexpectedly good episodes.

The 10 Most Unexpectedly Good Episodes

The Top 10 Unexpectedly Good episodes are:

# Define Elements for Manual Label
color = c("Actual\nRating" = "darkred",
          "Predicted\nRating" = 'black', 
          "Series\nAverage" = "darkblue")
shape = c("Actual\nRating" = 19, 
          "Predicted\nRating" = 19, 
          "Series\nAverage" = 1)

# Subset to only the Unexpectedly Good Results
good_results <- results %>% 
  filter(.resid > 0)  %>% 
  slice_max(order_by = remainder, n = 10, with_ties = F)

# Plot
good_results %>% 
  select(lbl, .resid, Predicted = .fitted, Actual = averageRating, 
         lower, upper, remainder) %>% 
  ggplot(aes(x = fct_reorder(lbl, remainder))) + 
  geom_pointrange(aes(y = Predicted, ymin = lower, ymax = upper, 
                      color = 'Predicted\nRating')) + 
  geom_point(aes(y = Actual, color = "Actual\nRating"), size = 2) +
  geom_text(aes(label = Actual, y = Actual), color = 'darkred', nudge_x = .3) +
  geom_text(aes(label = round(lower, 1), y = lower),  nudge_x = .3, size = 3) +
  geom_text(aes(label = round(upper, 1), y = upper),  nudge_x = .3, size = 3) +
  geom_text(aes(label = round(Predicted, 1), y = Predicted),  nudge_x = .3) +
  scale_color_manual(values = color, name = '') +
  labs(x = "", y = "IMDB Rating", 
       title = "Top 10 Unexpectedly <i style = 'color:#2E8B57'>Good Episodes </i>",
       subtitle = "*As measured by the difference between Prediction Interval and Actual IMDB Episode Rating*") +
  coord_flip() + 
  theme(
    plot.title = element_markdown(),
    plot.title.position = 'plot',
    plot.subtitle = element_markdown(size = 10),
    panel.grid.major.x = element_blank(),
    axis.text.x = element_markdown(size = 10),
    axis.title.x = element_text(size = 11),
    axis.text.y = element_markdown(size = 9),
    legend.position = 'top',
    legend.margin = margin(0, 0, -5, 0),
    legend.text = element_text(size = 9),
    legend.key.size = unit(0.2, "cm")
  )

According to this method the most unexpectedly good episode of TV since 1990 is from The Fresh Prince of Bel-Air’s 4th Season entitled “Papa’s Got a Brand New Excuse. Its a pretty entertaining show in general but this episode contains quite possible the most iconic scene from the show. Inquirer.com called the end of this episode “among the most tear-jerking in sitcom history”.

I’m not really familiar with many of the other episodes from this list, but at least from number 1 it seems like the method works well. Remember this isn’t looking for the best TV episodes but the best ones from a show you wouldn’t expect.

The 10 Most Unexpectedly Bad Episodes

The Top 10 Unexpectedly Bad episodes are:

# Subset Data to only Unexpectedly Bad Results
bad_results <- results %>% 
  filter(.resid < 0) %>% 
  slice_min(order_by = remainder, n = 10, with_ties = F)

# Plot
bad_results %>% 
  select(lbl, .resid, Predicted = .fitted, 
         Actual = averageRating, lower, upper, remainder) %>% 
  ggplot(aes(x = fct_reorder(lbl, -remainder))) + 
    geom_pointrange(aes(y = Predicted, ymin = lower, 
                        ymax = upper, color = 'Predicted\nRating')) + 
    geom_point(aes(y = Actual, color = "Actual\nRating"), size = 2) +
    geom_text(aes(label = Actual, y = Actual), 
              color = 'darkred', nudge_x = .3) +
    geom_text(aes(label = round(lower, 1), y = lower),  
              nudge_x = .3, size = 3) +
    geom_text(aes(label = round(upper, 1), y = upper),  
              nudge_x = .3, size = 3) +
    geom_text(aes(label = round(Predicted, 1), y = Predicted),  
              nudge_x = .3) +
    scale_color_manual(values = color, name = '') +
    labs(x = "", y = "IMDB Rating", 
         title = "Top 10 Unexpectedly <i style = 'color:#b22222'>Bad Episodes </i>",
         subtitle = "*As measured by the difference between Prediction Interval and Actual IMDB Episode Rating*") +
    coord_flip() + 
    theme(
      plot.title = element_markdown(),
      plot.title.position = 'plot',
      plot.subtitle = element_markdown(size = 10),
      panel.grid.major.x = element_blank(),
      axis.text.x = element_markdown(size = 10),
      axis.title.x = element_text(size = 11),
      axis.text.y = element_markdown(size = 9),
      legend.position = 'top',
      legend.margin = margin(0, 0, -5, 0),
      legend.text = element_text(size = 9),
      legend.key.size = unit(0.2, "cm")
    )

I have seen more of the bad list than I did the good list. While I don’t know much about later seasons of Riverdale, I can speak more about #5 which comes from Stranger Things’ 2nd Season. This episode involved a side-quest of one of the main characters meeting a family member that hasn’t appeared in any episode since.

Also, the Scrubs episode that appears “My Night to Remember” is the only clip show from the series.

In both of these examples, they’re unexpectedly bad because the shows in general are good (they have the two highest upper bounds of the Top 10) but these two episodes did nothing to advance the plot and were ultimately filler. Much like the Ted Lasso episode that motivated this analysis.

Moral of the story is that people don’t like filler episodes.

Drilling into a Few Shows

Just for fun I wrote a general purpose function inspired by the RatinGraph charts that will show any TV Series’s trend-lines and expected range as well as highlight any unexpected episodes.

plot_shows <- function(title){
  # Run the anomaly function on a single show and return all the episodes
  get_anomalies(
    all_tv_details %>% 
      filter(primaryTitle == title),
    onlyAnomalies = F
  ) %>% 
    mutate(primaryTitle = title) %>%
    ggplot(aes(x = episodeOverall, y = averageRating, color = seasonNumber)) + 
    # Plot the expected value range
    geom_ribbon(aes(ymin = lower, ymax = upper), fill = 'lightblue', 
                color = NA, alpha = .3) + 
    # Plot the overall trend line across the entire Series
    geom_smooth( se = F, method = 'lm', lty = 2, color = 'grey60') +
    # Plot the trendlines for each season
    geom_smooth(aes(group = seasonNumber), se = F, method = 'lm', 
                lty = 2, show.legend = F) +
    # Plot the actuals for each episode
    geom_point(alpha = .5) + 
    # Add annotations for any outliers
    geom_label_repel(data = . %>% filter(anomaly == T), size = 3, 
                     min.segment.length = 0, 
                     aes(label = glue("Season {seasonNumber} Episode {episodeNumber}
                                    {episodeTitle}
                                    Rating: {averageRating}")),
                     show.legend = F) + 
    guides(color = guide_legend(nrow = 1)) +
    labs(x = 'Episodes', y = 'IMDB Rating', title = title, color = 'Season:') + 
    theme(
      plot.title = element_text(family = 'Roboto'),
      legend.position = 'bottom',
    )
  
}

First let’s look at the show with the most unexpectedly good episode, The Fresh Prince of Bel-Air.

plot_shows('The Fresh Prince of Bel-Air')

For most of its episodes the IMDB ratings are a solid 7.5. There are some ups and some downs but nothing like the 9.7 rating that “Papa’s Got a Brand New Excuse” received.

On the negative side, let’s look at Scrubs:

plot_shows('Scrubs')

You can see that the model accounts for Season 9 being rated worse than 1-8 although not being unexpected since the whole season is poorly rating. Also, the clip episode “My Night To Remember” being far worse at 5.4 then the 8 rating that Seasons 1-8 usually had.

Finally, let’s look at Ted Lasso since it was the inspiration for this post:

plot_shows('Ted Lasso')

No major outliers. And the episode “Beard After Hours” (the lowest green dot) while lower than expected doesn’t meet the extreme criteria to be included here.