How crowd size at protests and riots is coded in ACLED data

On 20 July 2026, ACLED is changing how it labels crowd sizes for “Protest” and “Riot” events. Previously, ACLED recorded crowd sizes in the “Tags” column using the exact figure or phrasing as reported in the source, or applied “no report” where no crowd size was reported. With this update, ACLED introduces an automated crowd size label that groups the reported turnout into one of five standardized size buckets. This enables enhanced analysis of demonstration trends by turnout and frequency.

The automated label is formatted as “crowd size=X,” where X is one of five predicted bucket labels:

How is crowd size reported by sources and recorded by ACLED?

The size of demonstrations is a commonly overlooked aspect of reporting, and when it is reported, estimates can differ widely between sources. Reporting of crowd sizes also differs across regions in consistency and specificity, often containing size descriptions that are ambiguous about the exact number of participants. Previously, ACLED extracted the reported size description exactly as was written in the source report and reflected in the Notes column to allow users to classify these according to their own needs. With the updated crowd size labels, ACLED provides its own classification of crowd sizes. For users who prefer their own classifications, the event summary in the Notes column continues to provide more detailed descriptions of crowd sizes.

Why is ACLED adding new categories for crowd size?

ACLED’s approach to crowd size labeling deals with the ambiguity around exact sizes by classifying according to broader size buckets. This allows reasonable estimates of group sizes despite exact numbers not being available and makes the tens of thousands of unique crowd size figures in the sources easier to process and compare. The purpose is to provide an analytically relevant and consistent automated estimate of participation across “Protests” and “Riots” events, rather than to reproduce the precise figure reported in any single source. 

The appropriate use of this label is for the analysis of large numbers of demonstration events. At an individual event level, the Notes provide the most relevant size description.

How do the automated estimates work? 

The labels are predicted using a Bidirectional Encoder Representations from Transformers (BERT) model for natural language processing (NLP). The model is trained to interpret the crowd size information in the Notes column, along with its context, and to categorize that information into a specific bucket. 

Where the source and note contain no crowd size information, the event is labeled “crowd size=small” by default, which by ACLED’s estimate is the most likely size for such events. This decision is based on the assumption that larger protests are typically reported on in more detail and tend to have their sizes reported in the first place, and very small protests being much less frequent in general.

How accurate is this NLP-based labeling?

The model was evaluated on a large sample of events and achieved an overall accuracy of 98%. The model’s weighted F1-score, a measure which balances both the rate of correct positive classifications as well as missed ones, was also 98%. When the model does make mistakes, they are almost always to an adjacent bucket, such as “very small” to “small” rather than substantially different ones like “very small” to “massive.” 

Key changes

• The crowd size label format will be updated on 20 and 21 July 2026 according to the regional update schedule. Users who built their workflows around the previous free-text crowd size values should update their structure. 
• Crowd size labels are available for data from 1 January 2019 onward.
• Crowd size labels are removed for data before 1 January 2019 (less than 5% of events).

Updated period: All data from 1 January 2019 to present

Countries affected: All countries

Event types affected: Protests and Riots, including Mob violence