News Network Tone Analysis

Overall Tone Analysis

In this analysis, Fox News is generally regarded as a right-leaning outlet, MSNBC as left-leaning, and ABC News as a centrist source. These distinctions help contextualize differences in tone over time, as each network’s political orientation may influence how stories are framed and presented to their audiences.

Overview

This section examines long-term sentiment trends across three major U.S. news networks—Fox News, MSNBC, and ABC News—using tone scores derived from the Global Database of Events, Language, and Tone (GDELT) from 2015 through 2025. These scores quantify the overall tone of news articles on a continuous scale, offering insight into the emotional framing of events and issues over time.

Rather than centering on specific political events, this analysis takes a broad temporal view to uncover overarching trends in news tone. We investigate whether certain networks consistently portray the news with a more positive or negative tone, and how those patterns may shift across months and years.

Understanding GDELT Tone Scores

Before diving into the analysis, it’s essential to understand how GDELT tone scores are computed and what they represent. These scores provide a quantitative measure of emotional tone in global news coverage, enabling systematic comparisons across sources and time periods.

• Tone Score: This metric typically ranges from -10 (extremely negative) to +10 (extremely positive), with 0 indicating a neutral tone. It reflects the overall sentiment conveyed in a news article or segment.
• Calculation Method: GDELT applies natural language processing (NLP) techniques to extract sentiment by analyzing the frequency and intensity of positive and negative language within each document.
• Composite Measure: The tone score is derived as the difference between positive and negative sentiment components, providing a net emotional tone. In later sections, we’ll explore these components individually for a more detailed breakdown.

Higher tone scores indicate a stronger presence of positive language, while lower scores reflect more negative framing. These values allow us to track and visualize long-term sentiment trends, evaluate tone consistency or volatility, and compare differences in emotional framing across news networks with varying political orientations.

import pandas as pd
import glob
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
from collections import Counter
from scipy.stats import ttest_ind
import matplotlib.dates as mdates
from matplotlib.ticker import MaxNLocator

# Set visualization style
plt.style.use('seaborn-v0_8-whitegrid')
plt.rcParams['font.family'] = 'sans-serif'
plt.rcParams['font.sans-serif'] = ['Arial', 'DejaVu Sans', 'Liberation Sans']

# Define network colors
NETWORK_COLORS = {
    'Fox News': '#E41A1C',    # Red for Fox
    'MSNBC': '#377EB8',       # Blue for MSNBC
    'ABC News': '#984EA3'     # Purple for ABC
}

# Import data files
csv_files = (
    glob.glob("../data/fox/fox*.csv") +
    glob.glob("../data/abc/abc*.csv") +
    glob.glob("../data/msnbc/msnbc*.csv")
)

# Read CSVs with safe fallback and warn if fallback needed
dfs = []
for file in csv_files:
    try:
        dfs.append(pd.read_csv(file))
    except UnicodeDecodeError:
        print(f"⚠️  Warning: Unicode error in '{file}', reading with latin1 fallback.")
        dfs.append(pd.read_csv(file, encoding='latin1'))

df = pd.concat(dfs, ignore_index=True)

# Select relevant columns
columns_of_interest = [
    "parsed_date", "url", "headline_from_url",
    "V2Themes", "V2Locations", "V2Persons",
    "V2Organizations", "V2Tone"
]
df = df[columns_of_interest]

# Convert date and extract network information
df["parsed_date"] = pd.to_datetime(df["parsed_date"], errors="coerce").dt.tz_localize(None)

# Extract network source from URLs
def extract_network(url):
    try:
        url = url.lower()
        if 'fox' in url:
            return 'Fox News'
        elif 'abc' in url:
            return 'ABC News'
        elif 'msnbc' in url:
            return 'MSNBC'
        else:
            return 'Unknown'
    except AttributeError:
        return 'Unknown'

# Add network column
df['network'] = df['url'].apply(extract_network)

# Extract tone components
tone_split = df["V2Tone"].str.split(",", expand=True)
df["tone"] = pd.to_numeric(tone_split[0], errors="coerce")
df["positive_score"] = pd.to_numeric(tone_split[1], errors="coerce")
df["negative_score"] = pd.to_numeric(tone_split[2], errors="coerce")

# Create month and year columns for aggregation
df['month'] = df['parsed_date'].dt.to_period('M')
df['year'] = df['parsed_date'].dt.year
df['month_year'] = df['parsed_date'].dt.strftime('%Y-%m')

⚠️  Warning: Unicode error in '../data/fox/fox2020.csv', reading with latin1 fallback.

Dataset Overview

Note on Sample Sizes: The data shows a smaller sample for MSNBC compared to Fox News and ABC News. These differences reflect availability via GDELT’s API. This discrepancy should be considered when interpreting results, as it may impact the representativeness of trends for MSNBC.

# Reorder the article counts
ordered_networks = ['MSNBC', 'ABC News', 'Fox News']
article_counts = df['network'].value_counts().reindex(ordered_networks)

# Plot
plt.figure(figsize=(8, 4))
bars = plt.bar(article_counts.index, article_counts.values,
               color=[NETWORK_COLORS[network] for network in article_counts.index])

plt.title('Number of Articles by News Network', fontsize=14, fontweight='bold')
plt.xlabel('News Network')
plt.ylabel('Number of Articles')
plt.xticks(rotation=0)
plt.grid(axis='y', alpha=0.3)

# Add count labels on top of the bars
for bar in bars:
    height = bar.get_height()
    plt.text(bar.get_x() + bar.get_width()/2., height + 0.1,
             f'{int(height):,}', ha='center', va='bottom', fontweight='bold')

plt.tight_layout()
plt.show()

Tone Distribution Analysis

The dashed horizontal line at 0 represents a neutral tone, serving as a visual reference point to highlight how all three networks tend to lean toward negative sentiment in their reporting. This trend may reflect the nature of media content itself, where negative events often receive more attention and coverage due to their perceived newsworthiness.

All three news networks skew slightly negative in their average tone scores, consistent with prior findings that news coverage tends to focus more on conflict, controversy, and crisis. Fox News has an average tone score of –2.69, MSNBC averages –2.74, and ABC News is the most negative on average at –3.10.

While the overall shapes of the tone distributions are broadly similar, a few important distinctions emerge. ABC News, despite being considered a centrist outlet, exhibits a slightly more negative average tone and a wider distribution, indicating greater variability in emotional framing across its stories. This suggests that ABC may present a broader range of sentiment—from highly negative to moderately positive—compared to the other networks, which tend to cluster more tightly around their respective means.

Figure 1: Tone score distribution across news networks

Outlier Analysis of Tone Scores by Network

The bar chart above visualizes the number of tone score outliers—both negative and positive—for each news network, based on the 1.5×IQR rule.

• ABC News has the highest number of outliers overall, with 1,733 negative and 643 positive outliers. This aligns with the earlier observation of ABC’s wider tone distribution, suggesting a greater range in emotional framing.
• Fox News reports 1,142 negative and 449 positive outliers, placing it in the middle across both categories.
• MSNBC, notably, shows 1,393 negative and 548 positive outliers—despite having fewer total articles in the dataset compared to ABC and Fox. This indicates that MSNBC’s tone scores, while stemming from a smaller sample, exhibit a relatively high rate of extreme sentiment (especially on the negative end).

This pattern reinforces earlier findings that MSNBC’s tone distribution is highly skewed and variable, and that ABC News, though centrist in political alignment, features the most extreme tone scores overall. Outliers play a key role in revealing how each network diverges from neutral framing, offering insight into the intensity of sentiment conveyed over time.

import pandas as pd
import matplotlib.pyplot as plt

# Set network order
ordered_networks = ['MSNBC', 'ABC News', 'Fox News']
outlier_counts = []

# Calculate outliers for each network
for network in ordered_networks:
    scores = df[df['network'] == network]['tone'].dropna()
    q1 = scores.quantile(0.25)
    q3 = scores.quantile(0.75)
    iqr = q3 - q1
    lower_bound = q1 - 1.5 * iqr
    upper_bound = q3 + 1.5 * iqr
    
    num_negative = (scores < lower_bound).sum()
    num_positive = (scores > upper_bound).sum()
    
    outlier_counts.append({'network': network, 'type': 'Negative', 'count': num_negative})
    outlier_counts.append({'network': network, 'type': 'Positive', 'count': num_positive})

# Create DataFrame
outlier_df = pd.DataFrame(outlier_counts)

# Pivot data for easier stacking
pivot_df = outlier_df.pivot(index='network', columns='type', values='count').fillna(0)

# Plot
plt.figure(figsize=(8, 4))
bars_neg = plt.bar(pivot_df.index, pivot_df['Negative'], label='Negative', color='salmon', alpha=0.8)
bars_pos = plt.bar(pivot_df.index, pivot_df['Positive'], bottom=pivot_df['Negative'], label='Positive', color='skyblue', alpha=0.8)

# Add text labels
for i, network in enumerate(pivot_df.index):
    neg = pivot_df.loc[network, 'Negative']
    pos = pivot_df.loc[network, 'Positive']
    
    # Label for negative
    if neg > 0:
        plt.text(i, neg / 2, f'{int(neg)}', ha='center', va='center', fontsize=10, fontweight='bold', color='black')
    
    # Label for positive
    if pos > 0:
        plt.text(i, neg + pos / 2, f'{int(pos)}', ha='center', va='center', fontsize=10, fontweight='bold', color='black')

# Styling
plt.title('Number of Tone Score Outliers by News Network', fontsize=16, fontweight='bold')
plt.ylabel('Number of Outliers')
plt.xlabel('News Network')
plt.legend(title='Outlier Type')
plt.grid(axis='y', alpha=0.3)
plt.tight_layout()
plt.show()

Long-Term Tone Trends (2015-2025)

The analysis spans a full decade, capturing evolving sentiment during a wide range of historical events—including presidential election cycles, natural disasters, social movements, and public health crises. Rather than focusing solely on isolated events, this section prioritizes broad temporal trends to uncover patterns in how sentiment varies within and across networks.

Animated monthly average tone trends by news network (2015-2025)

The background shading marks changes in presidential administrations:

• Blue indicates Democratic leadership (Obama, Biden)
• Red represents Republican leadership (Trump’s terms)

Key insights include:

• Fox News maintains a relatively less negative tone, with a slight increase in sentiment during Trump’s presidencies.
• MSNBC exhibits sharper dips and greater volatility, particularly negative during both Trump terms, reflecting its more critical coverage.
• ABC News stays consistently negative but comparatively stable, suggesting a more neutral editorial stance.

Statistical Analysis

To test whether these tone differences are statistically meaningful, we conducted independent sample t-tests between each network pair.

• Fox News vs ABC News: A large t-statistic (25.91) and a p-value < 0.0001 indicate a highly significant difference in tone, with ABC News being significantly more negative.
• Fox News vs MSNBC: A smaller but still significant difference was found (p = 0.0031), suggesting Fox is consistently less negative than MSNBC.
• ABC News vs MSNBC: The negative t-statistic (-19.80) confirms ABC News is significantly more negative than MSNBC as well.

All comparisons yielded statistically significant results (p < 0.01), reinforcing that tone differences between these networks are not due to random chance but reflect meaningful editorial or coverage differences.

Statistical Significance Testing (t-test for tone differences):

Table 1: Statistical significance of tone differences between networks

	Comparison	t-statistic	p-value	Significant
0	Fox News vs ABC News	25.9065	0.0000	Yes
1	Fox News vs MSNBC	2.9581	0.0031	Yes
2	ABC News vs MSNBC	-19.8019	0.0000	Yes

Network Tone Comparison

Here, see that GDELT and AFINN produce relatively tight and centered score distributions when normalized to a 0–1 scale. In contrast, VADER shows a much wider range, suggesting it’s more sensitive to subtle tonal shifts.

import pandas as pd
import glob
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
from collections import Counter
from scipy.stats import ttest_ind
import matplotlib.dates as mdates
from matplotlib.ticker import MaxNLocator
from datetime import timedelta

# Set visualization style
plt.style.use('seaborn-v0_8-whitegrid')
plt.rcParams['font.family'] = 'sans-serif'
plt.rcParams['font.sans-serif'] = ['Arial', 'DejaVu Sans', 'Liberation Sans']

# Define network colors
NETWORK_COLORS = {
    'Fox News': '#E41A1C',    # Red for Fox
    'MSNBC': '#377EB8',       # Blue for MSNBC
    'ABC News': '#984EA3'     # Purple for ABC
}

# Import data files
csv_files = (
    glob.glob("../data/fox/fox*.csv") +
    glob.glob("../data/abc/abc*.csv") +
    glob.glob("../data/msnbc/msnbc*.csv")
)

# Read CSVs safely with fallback
dfs = []
for file in csv_files:
    try:
        dfs.append(pd.read_csv(file))
    except UnicodeDecodeError:
        dfs.append(pd.read_csv(file, encoding="latin1"))

df = pd.concat(dfs, ignore_index=True)

# Select relevant columns
columns_of_interest = [
    "parsed_date", "url", "headline_from_url",
    "V2Themes", "V2Locations", "V2Persons",
    "V2Organizations", "V2Tone",
    "afinn_tone_score", "vader_tone_score", "sentiment_label"
]

df = df[columns_of_interest]

# Convert date and extract network information
df["parsed_date"] = pd.to_datetime(df["parsed_date"], errors="coerce").dt.tz_localize(None)

# Extract network source from URLs
def extract_network(url):
    try:
        url = url.lower()
        if 'fox' in url:
            return 'Fox News'
        elif 'abc' in url:
            return 'ABC News'
        elif 'msnbc' in url:
            return 'MSNBC'
        else:
            return 'Unknown'
    except AttributeError:
        return 'Unknown'

# Add network column
df['network'] = df['url'].apply(extract_network)

# Extract tone components
tone_split = df["V2Tone"].str.split(",", expand=True)
df["tone"] = pd.to_numeric(tone_split[0], errors="coerce")
df["positive_score"] = pd.to_numeric(tone_split[1], errors="coerce")
df["negative_score"] = pd.to_numeric(tone_split[2], errors="coerce")

# Create month and year columns for aggregation
df['month'] = df['parsed_date'].dt.to_period('M')
df['year'] = df['parsed_date'].dt.year
df['month_year'] = df['parsed_date'].dt.strftime('%Y-%m')

# Define election dates
elections = {
    "2016": pd.to_datetime("2016-11-08"),
    "2020": pd.to_datetime("2020-11-03"),
    "2024": pd.to_datetime("2024-11-05")
}

# Add flag for period around each election
election_windows = []
for year, date in elections.items():
    df_sub = df[
        (df["parsed_date"] >= date - timedelta(days=30)) &
        (df["parsed_date"] <= date + timedelta(days=30))
    ].copy()
    df_sub["election_year"] = year
    df_sub["period"] = np.where(
        df_sub["parsed_date"] < date, "Before", "After"
    )
    election_windows.append(df_sub)

df_elections = pd.concat(election_windows)

df_elections = df_elections[[
    "parsed_date", "network", "election_year", "period",
    "tone", "afinn_tone_score", "vader_tone_score"
]]

# Function to normalize values based on theoretical ranges
def normalize_score_theoretical(series, min_val, max_val):
    return (series - min_val) / (max_val - min_val)

# Standard theoretical ranges for each sentiment measure
# GDELT Tone: typically ranges from -100 to +100
# AFINN: ranges from -5 to +5 per word, but articles can have wide ranges like -500 to +500
# VADER: ranges from -1 to +1

# Create normalized versions using theoretical ranges
df_elections['tone_normalized'] = normalize_score_theoretical(df_elections['tone'], -100, 100)
df_elections['afinn_normalized'] = normalize_score_theoretical(df_elections['afinn_tone_score'], -500, 500)
df_elections['vader_normalized'] = normalize_score_theoretical(df_elections['vader_tone_score'], -1, 1)

# Clip values to ensure they fall within 0-1 range (in case of outliers beyond theoretical ranges)
df_elections['tone_normalized'] = df_elections['tone_normalized'].clip(0, 1)
df_elections['afinn_normalized'] = df_elections['afinn_normalized'].clip(0, 1)
df_elections['vader_normalized'] = df_elections['vader_normalized'].clip(0, 1)

# Reshape the data for plotting with normalized scores
df_long_normalized = df_elections.melt(
    id_vars=["parsed_date", "network", "election_year", "period"],
    value_vars=["tone_normalized", "afinn_normalized", "vader_normalized"],
    var_name="model",
    value_name="score"
)

# Update the model names for better readability
df_long_normalized['model'] = df_long_normalized['model'].replace({
    'tone_normalized': 'GDELT Tone',
    'afinn_normalized': 'AFINN', 
    'vader_normalized': 'VADER'
})

# Set the visual style
sns.set(style="whitegrid", font_scale=1.1)

# Create the boxplot with normalized scores
plt.figure(figsize=(8, 5))
ax = sns.boxplot(
    x="model", 
    y="score", 
    data=df_long_normalized, 
    palette="Set2", 
    showfliers=False
)

# Add original scale information as text annotation
plt.figtext(
    0.01, 0.01, 
    "Original scales - GDELT: [-100, 100], AFINN: [-500, 500], VADER: [-1, 1]",
    fontsize=9
)

plt.title("Normalized Sentiment Score Comparison (All Elections & Networks)", fontsize=14)
plt.xlabel("Sentiment Model")
plt.ylabel("Normalized Score (0-1 scale)")
plt.tight_layout()
plt.show()

It’s also worth noting that GDELT and AFINN seem to have similar distributions despite using different scales and methodologies, which suggests they might be capturing similar sentiment patterns in the news articles.

GDELT vs. AFFIN

We directly compare raw GDELT and AFINN scores. Each point is an article. The positive correlation of 0.42 shows that while the models often agree, there are many cases where they don’t — especially when GDELT classifies something as negative and AFINN still reads it as positive. Again, this reiterates how unreliable these tools can be.

import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5))

correlation = df_elections['tone'].corr(df_elections['afinn_tone_score'])

scatter = plt.scatter(
    df_elections['tone'], 
    df_elections['afinn_tone_score'],
    alpha=0.5,
    color='red' 
)

# Add correlation information
plt.annotate(
    f'Correlation: {correlation:.3f}', 
    xy=(0.05, 0.95), 
    xycoords='axes fraction', 
    fontsize=8,
    bbox=dict(boxstyle="round,pad=0.3", fc="white", ec="gray", alpha=0.8)
)

# Add a horizontal and vertical line at 0 to indicate neutral sentiment
plt.axhline(y=0, color='blue', linestyle='--', alpha=0.7)
plt.axvline(x=0, color='blue', linestyle='--', alpha=0.7)

# Label the quadrants
plt.text(df_elections['tone'].max()*0.7, df_elections['afinn_tone_score'].max()*0.7, 
         'Both Positive', fontsize=10, ha='center')
plt.text(df_elections['tone'].min()*0.7, df_elections['afinn_tone_score'].max()*0.7, 
         'GDELT Negative\nAFINN Positive', fontsize=8, ha='center')
plt.text(df_elections['tone'].max()*0.7, df_elections['afinn_tone_score'].min()*0.7, 
         'GDELT Positive\nAFINN Negative', fontsize=8, ha='center')
plt.text(df_elections['tone'].min()*0.7, df_elections['afinn_tone_score'].min()*0.7, 
         'Both Negative', fontsize=10, ha='center')

# Add labels and title
plt.xlabel('GDELT Tone Score')
plt.ylabel('AFINN Tone Score')
plt.title('GDELT vs. AFINN Sentiment Comparison', fontsize=10)

# Add a grid for better readability
plt.grid(True, alpha=0.3)

# Tight layout
plt.tight_layout()

# Show the plot
plt.show()

A significant cluster of points falls in the “Both Negative” quadrant, confirming your observation that both GDELT and AFINN detect substantial negative sentiment in the news articles.

There are very few points in the “GDELT Positive, AFINN Negative” quadrant, suggesting that when GDELT finds positive sentiment, AFINN rarely strongly disagrees.

However, there are a fair number of points in the “GDELT Negative, AFINN Positive” quadrant, indicating some systematic differences in how these two metrics evaluate certain types of content.

Conclusion

This analysis of tone scores across Fox News, MSNBC, and ABC News from 2015 to 2025 reveals several important patterns in how major U.S. news networks emotionally frame their coverage. Our findings provide empirical evidence for both common assumptions about media bias and more nuanced insights about how news sentiment varies over time.

Key Findings

1. Persistent Negative Bias: All three networks consistently maintain negative average tone scores throughout the decade, supporting the long-established media principle that “if it bleeds, it leads.” This industry-wide tendency to emphasize negative stories reflects both commercial incentives and journalistic norms that prioritize conflict, crisis, and controversy.

2. Network-Specific Patterns: Despite the overall negative trend, statistically significant differences emerged between networks:

   • Fox News maintains the least negative tone on average (-2.69), with notable upticks during Republican administrations
   • MSNBC shows greater volatility (-2.74 average), with pronounced negative spikes during Trump’s terms
   • ABC News, despite its reputation for centrism, displays the most consistently negative tone (-3.10) with the widest distribution of scores

3. Political Alignment Effects: The background shading highlighting presidential administrations reveals clear patterns where network tone often aligns with political affiliation. Fox News sentiment improves during Republican leadership, while MSNBC sentiment dips more sharply during these periods, suggesting a substantial relationship between political alignment and emotional framing.

4. Outlier Analysis: The significant number of outliers, particularly from ABC News and MSNBC despite its smaller sample size, indicates that extreme framing—especially negative framing—plays an important role in how these networks cover certain stories.