University of Twente Student Theses


Analyzing the accuracy of input information to the "last time buy" decision process at IBM SPO

Kneppers, Erik (2013) Analyzing the accuracy of input information to the "last time buy" decision process at IBM SPO.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Abstract:Introduction and goals – This thesis describes a research project on LTB decisions at IBM SPO. Outcome of an LTB decision is a final buy quantity of a spare part, needed to cover demand for that part during the remaining service period. The LTB decision is based on a calculation involving several input parameters, such as current monthly demand for a part and the decline pattern of demand. Most parameters consist of forecasted values and errors made in these forecasts result in the final buy quantity being too high or too low. Previous research concluded that safety quantities were added for several parameters and at several steps in the decision process leading to overestimated quantities. A recommendation was to only add a safety quantity at the end. It was concluded that it is hard to determine how large that safety factor should be as the IBM LTB model is completely deterministic and does not account for errors in input values; the outcome of the decision process is a single value. This value is however uncertain and the characteristics of uncertainty are currently not known. Introducing a stochastic model requires knowledge of the uncertainty of input information: the accuracy of forecasted values. Measuring this is also useful to identify where improvements can be made or where the process can be simplified. For these reasons, measuring accuracy of forecasted input parameters was the main goal of this research project. A focus was placed on forecasts of the decline pattern as this is a very important input parameter and (so far) an underexposed subject. Methodology – The methods behind the prediction of input parameters were described and decomposed, identifying sources of errors. Measuring these errors was done by data analysis using the IBM SPSS Modeler software package. The scope of the research was limited to LTBs for parts in the Power hardware division. Bias (average error) for systematic errors and mean absolute percentage error (MAPE) were calculated for forecasted input. Missing data was a recurrent problem. Forecasted and actual installed base information was found to be incomplete and several methods in SPSS Modeler were used to circumvent this. Also, analysis was complicated by the fact that data had to be obtained from different sources (databases, Lotus Notes, Excel). Important to note about the methodology is that uncertainty characteristics of parameters for a single part were estimated by measuring average error characteristic over all (Power LTB) parts. If specific information is available for a part, it might be possible to provide better estimates with regard to uncertainty but in the absence of such information the average values can be used. Results and conclusions – With regards to decline factors, several potential sources of errors were found. Most obvious is the accuracy of installed base forecasts of machine models. Part of the errors in these forecasts can be explained by service extensions that took place after the LTB decision was made. However, even if installed base forecast per machine are perfectly accurate, the decline pattern being accurate also depends on the assumption that the trend in installed base (of the selected machines) perfectly predicts the trend in usage. It was found that installed base forecasts are systematically overestimated by 20% of the forecasted value. Variability of the error, after correcting for this bias, amounted to a MAPE of around 30% for the cumulative sum of installs for all years. As a part is often used in multiple machines/models, variability of the decline pattern is lower in practice. Higher commonality of a part results in lower variability, with MAPE in normal cases being between 15 and 20%. The assumptions that installs predict usage was tested by predicting usage based on actual installed base numbers related to installed base and usage in the first year. It was found that the ‘forecasts’ that are obtained in that manner are unbiased and measure, in normal cases, a MAPE below 10%. It can be concluded that the method itself works: the actual trend in installed base is a good predictor for the trend in usage. Errors in base usage forecasts are another large factor in errors in LTB quantities. These forecasts were found to be biased by 9% of the actual value and MAPE was measured at around 40% for slow movers with yearly demand below 10, 20% for yearly demand between 10 and 100 and below 10% for fast movers. While the low accuracy for slow-movers is to be expected, it raises the question if the method of using monthly forecast as input parameter is a suitable approach in if it should not be replaced with the previous year of actual demand. Errors in predicted return rate and repair yield were hard to measure as the values used for LTB calculations could not easily be obtained. However, it was found that return rate and repair yield differ between parts and over time: it should be set to a single average value and used in a deterministic model. Recommendations – On of the main recommendations is to better log forecasted and actual installed base data. These can then be compared, accuracy in terms of bias and MAPE can be calculated and feedback should be sent to Service Planning at least once a year, enabling this department to improve their forecasts. The SPSS Modeler models can be used for this. Logging of estimated return rate and repair yield should be fixed by the introduction of the PLCM support application. This application also uses a somewhat different forecasting method for the base usage and these new forecasts should be analyzed to compare performance against the Xelus forecasts. The accuracy of forecasted return rate and repair yield should be measured to better estimate uncertainty in the LTB quantity. Goals for improvement should be set based on the accuracy of parameters rather than just the overall performance in terms of scrapped parts / parts short. In the long term, when measuring of parameter forecast accuracy has been fully implemented, SPO should replace the current deterministic model with a stochastic model to enable explicitly selecting a level of risk.
Item Type:Essay (Master)
Faculty:BMS: Behavioural, Management and Social Sciences
Subject:85 business administration, organizational science
Programme:Business Administration MSc (60644)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page