"background": "The management of industrial machinery fleets in developing economies is often hampered by a lack of robust, data-driven tools for forecasting operational costs and asset performance. In Tanzania, this leads to suboptimal capital allocation and maintenance scheduling, reducing the cost-effectiveness critical for industrial development. ", "purpose and objectives": "This paper presents a methodological evaluation of a novel time-series forecasting model designed to measure and predict the cost-effectiveness of heavy machinery fleets. The objective is to assess the model's predictive accuracy and operational utility within the Tanzanian industrial context. ", "methodology": "The proposed model integrates Autoregressive Integrated Moving Average (ARIMA) components with exogenous maintenance and utilisation variables. The core forecasting equation is Ct = \ + =1^{p\ Ct-i + =1^q\ -j + =1^m\ Xk, t + \, where Ct represents cost per operating hour. Model parameters were estimated using maximum likelihood, and 95% confidence intervals were generated for all forecasts. Evaluation was conducted using a rolling-origin forecast evaluation on a proprietary dataset of fleet operations. ", "findings": "The model demonstrated a statistically significant reduction in forecast error compared to a naive seasonal benchmark, with a mean absolute percentage error (MAPE) of 8. 7% (95% CI: 7. 2% to 10. 1%). A key finding was that incorporating planned maintenance schedules as an exogenous variable explained approximately 22% of the variance in unexpected downtime costs. ", "conclusion": "The evaluated time-series model provides a statistically sound and operationally relevant method for forecasting machinery fleet cost-effectiveness. It offers a substantial improvement over simpler forecasting techniques commonly employed in the region. ", "recommendations": "Fleet managers should adopt integrated forecasting models that combine intrinsic cost time-series with planned maintenance data. Further research should focus on validating the model across
Mwinyi et al. (2021) studied this question.