Generate Streaming Platform Synthetic Data in Python

Streaming platform data has a critical coherence requirement: churned_at must be null for active subscribers and non-null for churned ones, and is_churned must match. Get this wrong and your churn prediction model trains on contaminated labels. Misata generates a four-table streaming dataset where is_churned and churned_at are always consistent, content types follow realistic catalog distributions (series 55%, movies 35%, documentaries 10%), and watch completion rates match real streaming benchmarks (~65% for movies).

import misata

tables = misata.generate(
    "A Netflix-like streaming service with 10k subscribers and a content library",
    rows=10_000,
    seed=42,
)
print(list(tables.keys()))   # ['subscribers', 'content', 'watch_history', 'ratings']

# churned_at is null for all active subscribers
active = tables["subscribers"][~tables["subscribers"]["is_churned"]]
assert active["churned_at"].isna().all()

What Misata generates#

Four tables: subscribers, content, watch_history (linking subscribers to content), and ratings. Churn logic is coherent, content distribution is realistic, and viewing patterns match real streaming behavior.

Tables and columns#

Realistic distributions#

• churned_at is null for active subscribers: is_churned and churned_at are always consistent
• Content type: series 55%, movies 35%, documentaries 10%, matching real platform catalog ratios
• Watch completion rate ~65% for movies; lower per-episode completion for series
• Plan distribution across basic, standard, and premium tiers with realistic uptake ratios
• Device mix across mobile, smart TV, desktop, and tablet, matching real streaming device splits

Quick start#

import misata
import pandas as pd

tables = misata.generate(
    "A streaming service with 5k subscribers and a diverse content library",
    rows=5000,
    seed=42,
)

# Churn rate
churn_rate = tables["subscribers"]["is_churned"].mean()
print(f"Churn rate: {churn_rate:.1%}")

# Watch completion by content type
merged = tables["watch_history"].merge(
    tables["content"][["content_id", "type"]], on="content_id"
)
print(merged.groupby("type")["completed"].mean())

# Plan distribution
print(tables["subscribers"]["plan"].value_counts(normalize=True))

Common use cases#

• Recommendation model training: use watch_history and ratings as the interaction matrix for collaborative filtering models before you have real viewing data
• Churn prediction pipelines: train models on subscriber engagement patterns (watch frequency, completion rates, last active date) against is_churned labels
• Content performance analytics: build watch time, completion rate, and rating dashboards by genre, type, and release year
• A/B test framework validation: generate subscriber cohorts with varied plan and country distributions to test experiment assignment pipelines
• Personalization engine testing: validate recommendation ranking logic and fallback strategies against a full content catalog
• Subscriber lifecycle QA: test onboarding, upgrade, downgrade, and cancellation workflows against subscribers with realistic join and churn patterns

Advanced: viral growth narrative#

tables = misata.generate(
    "Streaming service that gained subscribers rapidly after a viral original series, "
    "now facing increasing churn as the content catalog ages",
    rows=10_000,
    seed=42,
)

Advanced: locale-aware generation#

# Latin American streaming — Spanish and Portuguese content, regional subscriber distribution
tables = misata.generate("Latin American streaming platform with Spanish content library", rows=5000)

# Asian streaming — Korean, Japanese, and Chinese content types
tables = misata.generate("Asian streaming platform with K-drama and anime content", rows=5000)

Advanced: quality-guaranteed generation#

tables = misata.generate(
    "Streaming platform with 10k subscribers",
    min_quality_score=85,
    smart_correlations=True,  # auto-adds churn risk↔watch_duration correlation
    rows=10_000,
    seed=42,
)

Related guides#

• Multi-table Synthetic Data
• Narrative Patterns
• Database Seeding in Python
• Faker vs SDV vs Misata