Moment-to-Moment Engagement Prediction through the Eyes of the Observer: PUBG Streaming on Twitch

Key insights

5 takeaways

01
Artificial neural networks can predict live-stream viewer engagement with up to 84% accuracy by correlating in-game telemetry with Twitch chat frequency.See it on page 1
02
Classifying streamers into three distinct archetypes—Noob, Explorer, and Pro—using hierarchical clustering significantly improves the generalizability of engagement models across different broadcasting styles.See it on page 6
03
The predictive framework utilizes 40 distinct telemetry features from PUBG to map how specific player behaviors and competitive events drive audience interest.See it on page 1
04
Models trained on individual streamers are highly accurate but difficult to scale, necessitating the use of uncertainty bounds to filter ambiguous data and improve model consistency.See it on page 5
05
The research establishes that third-person observation of digital actions is a reliable foundation for quantifying real-time audience satisfaction without requiring streamer-specific training data.See it on page 3

Summary

Predicting moment-to-moment viewer engagement in live-streamed gaming relies on the correlation between in-game telemetry and audience activity. By utilizing Twitch chat frequency as a reliable proxy for viewer interest, artificial neural networks can effectively forecast engagement levels based on specific gameplay events. This approach demonstrates that third-person observation of digital actions provides a robust foundation for understanding audience dynamics, achieving prediction accuracies of up to 84% when models are tailored to specific gameplay contexts.

The research focuses on the battle royale title PUBG, analyzing 40 distinct telemetry features to map how player behavior influences viewer response. While models trained on individual streamers yield high accuracy, scaling these predictions across diverse broadcasting styles presents greater complexity. To address this, the methodology incorporates uncertainty bounds to filter ambiguous data and employs hierarchical clustering to categorize streamers into three distinct archetypes: Noob, Explorer, and Pro. This classification strategy significantly improves model generalizability, allowing for consistent performance across varied play styles without requiring streamer-specific training.

These findings establish a scalable framework for developers and streaming platforms to quantify audience engagement in real time. By moving beyond individual streamer metrics toward a style-based predictive model, the research offers a sophisticated tool for analyzing the intersection of player performance and viewer satisfaction. Ultimately, the ability to predict engagement through telemetry alone highlights the deep connection between competitive gameplay events and the emotional investment of the streaming audience, providing a foundation for future developments in affective computing and interactive media design.