Software Forecasting Models

Abstract

Several forecasting models were evaluated with respect to their forecasting accuracy using NASA Space Shuttle failure data. Simple models like exponential smoothing are attractive if the forecasting range can be limited. This is the case because these models use the most recent data in making a forecast. In other cases, where forecasts must be made over a wide range, software reliability models are superior. The reason is that these models use a larger set of data in making parameter estimates. This feature results in more accurate forecasts over a wider range of time periods in the future. We found that significant improvements could be made to forecasting accuracy for all models by the simple process of modifying original forecasts based on the relative errors of those forecasts.