This model predicts the call center metrics for August 2026 across four distinct portfolios (A, B, C, and D) to optimize staffing. Accurate forecasting of Call Volume (CV), Average Handle Time (CTT), and Abandon Rates (ABD) is important for workforce management. By predicting these metrics for August 2026, this tool provides the organization the ability to optimize staffing levels, ensuring an adeqate number of agents are available to maintain service levels but stick to budget targets.
Our model relies on a mix of categorical and temporal features to account for the seasonal nature of call center traffic. We used temporal features like month, day of the month, and day of the week to find the standard call patterns. We used US public holidays to account for significant drops in the pattern. To account for the cycle of weeks, we appy cyclical encoding to time variables using sine and cosine transformations. Additionally, we utilize historical lag variables, such as volume from the same day in the previous year, to provide a consistent baseline for the regressor and improve overall predictive stability.
We utilize a top-down approach to handle interval-level high-variance, ensuring our 30-minute forcasts are accurate to daily totals. We start with stage 1, where individual XGBoost Regressors are trained for each portfolio to predict daily totals for call volume and handling time. In stage 2, these daily aggregates are transitioned into 30-minute intervals using HIstorical Intraday Profiles created by grouping past data by day of week and IntervalIdx. By calculating the specific metrics for each interval, and multiplying it by the stage 1 daily prediction, we generate a final 48-interval forecast that is guaranteed to sum back to the predicted daily total.
The model achieved a strong predictive accuracy across all four portfolios, particularly for Call Volume and CCT, with Mean Absolute Percentage Errors (MAPE), ranging from 9.5% and 11.5% for volume and 1.9% to 2.6% for handling times. While the Abandon Rate shows a higher MAPE exceeding 100%, this is mainly caused by the high frequency of zero-value intervals in the historical data, whihc disproportionately inflates percentage-based error metrics. Overall, the stage 1 daily forecasts provided a reliable baseline that, when combined with the stage 2 intraday profiles, resulted in a mathematically accurate 30-minute forecast for the August 2026 period.