How it works
ML attendance predictions are generated using a machine learning model that identifies patterns in past attendance and applies them to future dates.
The model incorporates:
Historical attendance data (badge data)
Captures daily attendance patterns over time. A minimum of ~12 months of data is required.Historical headcount data
Helps the model better understand attendance patterns and improve prediction accuracy over time.Weather data
Includes signals like temperature, rain, and severe weather. This is location-specific to each building and is included because weather patterns are highly correlated with changes in attendance.Holiday calendars (when provided)
Region-specific holidays are included when uploaded. These help account for expected drops in attendance.Event data (when provided)
Office or company events can influence attendance patterns. This data must be provided manually and is not always included.
Predictions are continuously updated as the model processes new data and improves accuracy over time.
What the model outputs
The model generates daily attendance predictions at the building level, which are then aggregated for use in Portfolio Planner.
Based on badge activity, not utilization or seat demand
Shown as a forecast line in the supply and demand graph in Portfolio Planner
Can be viewed as:
Average predicted attendance
A smoothed view of expected attendance that reduces the impact of unusually high or low days.Peak predicted attendance
Based on the highest expected attendance levels, often appearing as a flatter line since it reflects peak demand.P80 predicted attendance
Represents the 80th percentile of expected attendance, capturing typical high-demand levels while excluding extreme outliers.
What this enables
ML attendance predictions support near-term planning decisions such as:
Staffing and services
Plan food and beverage, facilities, and on-site support based on expected attendance.Transportation planning
Adjust shuttle schedules and capacity based on projected demand.Space planning
Anticipate when buildings may be more or less utilized.
Example:
A one-month lookahead shows lower predicted attendance on Fridays. Teams can reduce food orders and shuttle frequency to avoid unnecessary cost and waste.
Data requirements
To generate reliable predictions, the model requires:
~12 months of historical attendance data (required)
Historical headcount data (required)
Holiday calendar (optional, when provided)
Event data (optional, when available)
Weather data is included automatically.
Accuracy and time horizon
In testing, predictions up to one week out have shown 95%+ accuracy
Predictions are most reliable in the near term
As the planning horizon extends, predictions should be used as directional guidance rather than precise estimates.
Things to keep in mind
Predictions improve with more historical data.
New buildings with limited history will have lower accuracy.
The model may take time to adjust to sudden changes (e.g., policy shifts, hiring changes), typically requiring a short period of new data to reflect updated trends.
Attendance predictions are primarily based on badge data, with additional inputs like historical headcount used to improve accuracy. They do not reflect assigned seats or theoretical capacity.
Event data can improve accuracy when available, but is not always included.