Contact center workload forecasts covering varying timeseries granularities and operating horizons

A computer-implemented method for selecting forecasting models for generating improved timeseries workload forecasts for a contact center covering varying timeseries granularities and operating horizons, the method comprising the steps of:

• selecting, via a first selection process, a first select forecasting model from first candidate forecasting models for forecasting a workload level in accordance with a lower-granularity timeseries, wherein the first selection process comprises the steps of:
• receiving a first timeseries dataset, the first timeseries dataset comprising workload data recorded during a past operating period of the contact center;
• defining different timeseries datasets within the first timeseries dataset, the different timeseries datasets including at least a first shorterm dataset and a first longterm dataset;
• testing, via a machine-learning model validation, using the first longterm dataset, each of the first candidate forecasting models in accordance with a first longterm cross-validation process that includes multiple folds, wherein the first longterm cross-validation process includes modifying a hyper-parameter of and retraining with machine learning and based on the modified hyper-parameter the each first candidate forecasting model for each fold of the multiple folds of the longterm cross-validation process and calculating a longterm accuracy score for each of the first candidate forecasting models;
• testing, via the machine-learning model validation, using the first shorterm dataset, each of the first candidate forecasting models in accordance with a first shorterm cross-validation process that includes multiple folds, wherein the first shorterm cross-validation process includes modifying a hyper-parameter of and retraining with machine learning and based on the modified hyper-parameter the each first candidate forecasting model for each fold of the multiple folds of the shorterm cross-validation process and calculating a shorterm accuracy score for each of the first candidate forecasting models;
• calculating, for each of the first candidate forecasting models, a combined accuracy score that comprises a combination of the longterm accuracy score and the shorterm accuracy score of the first longterm cross-validation process and the first shorterm cross-validation process, respectively;
• comparing the combined accuracy scores for each of the first candidate forecasting models and, based on the comparison, selecting one of the first candidate forecasting models as the first select forecasting model;
• identifying a second select forecasting model, wherein the second select forecasting model is configured to forecast a distribution that distributes a workload level across a higher-granularity timeseries defined within the lower-granularity timeseries;
• automatically, in response to identifying the first select forecasting model and the second select forecasting model:
• executing at least one instance of the first select forecasting model to forecast a workload level applicable to the lower-granularity timeseries for a future operating period of the contact center;
• and
• executing, by providing the forecasted workload level for the future operating period as input thereto, at least one instance of the second select forecasting model to forecast a distribution that distributes the forecasted workload level of the future operating period across the higher-granularity timeseries defined therewithin;
• combining the forecasted workload level forecasted by the first select forecasting model and the distribution forecasted by the second select forecasting model to produce the continuous timeseries workload forecasts for the contact center applicable to the varying timeseries granularities;
• wherein the second select forecasting model is selected by a second selection process that comprises the steps of:
• receiving a second timeseries dataset, the second timeseries dataset comprising workload distribution data recorded during a past operating period of the contact center;
• defining different timeseries datasets within the second timeseries dataset, the different timeseries datasets including at least a second shorterm dataset and a second longterm dataset;
• testing, via the machine-learning model validation, using the second longterm dataset, each of the second candidate forecasting models in accordance with a second longterm cross-validation process, wherein the second longterm cross-validation process includes modifying a hyper-parameter of and retraining with machine learning and based on the modified hyper-parameter the each second candidate forecasting model for each fold of the multiple folds of the longterm cross-validation process and calculating a longterm accuracy score for each of the second candidate forecasting models;
• testing, via the machine-learning model validation, using the second shorterm dataset, each of the second candidate forecasting models in accordance with a second shorterm cross-validation process, wherein the second shorterm cross-validation process includes modifying a hyper-parameter of and retraining with machine learning and based on the modified hyper-parameter the each second candidate forecasting model for each fold of the multiple folds of the longterm cross-validation process and calculating a shorterm accuracy score for each of the second candidate forecasting models;
• calculating, for each of the second candidate forecasting models, a combined accuracy score that comprises a combination of the longterm accuracy score and the shorterm accuracy score of the second longterm cross-validation process and the second shorterm cross-validation process, respectively;
• comparing the combined accuracy scores for each of the second candidate forecasting models and, based on the comparison, selecting one of the second candidate forecasting models as the second select forecasting model;
• wherein, in the first selection process, the first timeseries dataset comprises between 3 and 5 years of timeseries data, and in the second selection process, the second timeseries dataset comprises between 0.5 and 1.5 years of timeseries data;
• wherein:
• the first longterm dataset is defined as including substantially an entirety of the first timeseries dataset;
• and
• the first shorterm dataset is defined as a continuous segment terminating at a most recent end of the first timeseries dataset that encompasses between 35%-50% of the first timeseries dataset;
• wherein:
• the second longterm dataset is defined as including substantially an entirety of the second timeseries dataset;
• and
• the second shorterm dataset is defined as a continuous segment terminating at a most recent end of the second timeseries dataset that encompasses between 15%-35% of the second timeseries dataset,
• wherein, in the first selection process, the first longterm cross-validation process and the first shorterm cross-validation process each comprises a cross-validation process that defines a training dataset and a test dataset according to a rolling horizon
• wherein, in the second selection process, the second longterm cross-validation process and the second shorterm cross-validation process each comprises a cross-validation process that defines a training dataset and a test dataset according to a rolling horizon;
• and
• fine tuning the hyper-parameters of the machine learning based on the cross-validation processes.