The problem owner of the present study is a consulting company that provides simulation-based workforce planning advice to a big manufacturing firm XYZ. The latter pertains the number of engineers of various skill levels that are needed for the repair of health care equipment in hospitals of a large country. The prediction of machine failures (reliability forecasting) is a crucial input to the simulations that affects the quality of the business advice. Currently, the problem owner follows a reliability forecasting approach based on lifetime models following the HPP [1]. Nevertheless, this practice has several limitations as: i) the predictive performance is not always satisfactory due to data overfitting (Liang, 2011), ii) real-world systems do not generally comply with the HPP traits (Kurien, Sekhon & Chawla, 1993), namely constant failure rates of a memoryless failure process, while reliability is non-linear and complex due to a bunch of factors (Chatterjee & Bandopadhyay, 2012). In the view of the above, the problem owner needs to increase the efficiency of workforce planning that will finally lead to cost savings for firm XYZ. It is believed that a more efficient planning can be achieved through the improvement of the forecasting approach. Forecasting should fulfil certain requirements, namely it should predict the failure patterns of multiple machines, at an acceptable level of accuracy, with a high degree of automation. Thus, the study’s research objective is defined as: to provide an automated forecasting framework that detects and predicts the failure patterns of multiple machines with acceptable accuracy. For achieving the research objective, firstly, a clarification of the forecasting requirements is done through a semi-structured interview with the problem owner. Among others, it is clarified that accuracy is the hourly absolute deviation between the actual and the forecasted inter-failure time of a machine (MAE), and it concerns only its next failure (one-step ahead forecasting). Additionally, for a bunch of reasons, two different levels of acceptable accuracy are defined, the 1st with minimum accuracy of 120h (1 working week), and the 2nd of 2160h (1 quarter). Secondly, the identification of the most promising forecasting approach that can fulfil the given requirements is done through a literature review. Time series forecasting is found to be the most promising approach as it: i) outperforms reliability models that follow the NHPP in terms of accuracy (Ho & Xie, 1998; Dindarloo & Siami-Irdermoosa, 2015; Fan & Fan, 2015), ii) is able for automated and large-scale application (Wagner et al., 2011). Subsequently, a case study, which pertains reliability forecasting of radiation treatment machines maintained by firm XYZ, is conducted in order to evaluate the time series approach. The reliability metric of Time-Between-Failures (TBF) is used for forecasting, whilst the time series cross-validation method is employed for its evaluation. The time series approach followed is based on the use of three parametric methods (ARIMA, Exponential Smoothing, Optimized Theta) and two artificial neural networks (FFNN, RGMDH) applied on the machine group level (2 groups) and on the individual machine level (5 machines). In this context, experimentations take place under both full and adjusted for outliers data conditions. Moreover, the related repair data, expressed by Time-To-Repair (TTR) and by a dummy variable that represents the use of spare items, is used in the TBF forecasting with ARIMAX models. The case study demonstrates that: i) on the machine group level, the series are white noise involving that the failure process is memoryless and failure patterns cannot be detected, ii) on the individual machine level, the best performing forecast model of every machine examined satisfies the 2nd level of acceptable accuracy. The MAE error metric of the best forecast model for each machine examined is substantially less than 2160h. Thus, the present study has succeeded in its objective. The reliability forecasting framework produced constitutes a holistic approach to the prediction of machine failures, as with its various and at a degree, complementary methods can deal with all the basic types of failure data (e.g. autocorrelations, seasonality, trend, non/linearity, etc.) The framework formed is provided to the problem owner allowing for the transformation of the workforce planning of firm XYZ from an annual to a quarterly basis. The recommendations for the problem owner as well as for future research are: first, the execution of experimental simulations with a planning horizon of 3 months in order to evaluate the possible cost savings for firm XYZ. Second, the collection of new relevant to machine failures data (e.g. machine utilization, purchase date), and third, the extension and evaluation of the forecasting framework with the inclusion of these new data and/or with new methods (e.g. hybrid, FFNN with external covariates) and techniques (e.g. time series clustering). Fourth, the application and re-evaluation of the reliability forecasting framework formed when the failure data of 2016 become available. Fifth, the use of failure behavior’s variability as a stakeholder management tool when the problem owner deals with forecasting projects. Last, the use of the time series cross-validation method for the evaluation of forecast models and the great amount of attention on the potential existence of outliers in the dataset. On reflection, the contribution of the present thesis is multi-dimensional. First, a holistic and multi-method reliability forecasting framework that can deal with almost any failure process has been produced. This framework can be used in relevant projects as it can be extended and adjusted to the conditions of each project. Second, the aforementioned framework has been built though a state-of-the-art analytical forecasting process that can also be used by the problem owner in different projects. Third, there is a clear potential for cost savings for firm XYZ if workforce planning is adjusted in a quarterly horizon. Fourth, there is a knowledge contribution to the performance of various time series methods (e.g. Optimized Theta, RGMDH) in the context of reliability forecasting. Fifth, there is a clear contribution to the increase of the domain knowledge of reliability forecasting in health care equipment in general, and in radiation treatment machines in particular. Last, it has been highlighted that the initial evaluation of the variability of the failure behavior of a set of machines can serve as a stakeholder management tool as regards the final forecasting deliverable.