Forecasting technology costs via the experience curve — Myth or magic?

Abstract

To further understand the effectiveness of experience curves to forecast technology costs, a statistical analysis using historical data is carried out. Three hypotheses are tested using available datasets that together shed light on the historical ability of experience curves to forecast technology costs. The results indicate that the Single Factor Experience Curve is a useful forecasting model when errors are viewed in their log format. Practitioners should note that due to the convexity of the log curve a mean overestimation of potential cost reductions can arise as values are converted into monetary units. Time is also tested as an explanatory variable, however forecasts made with endogenous learning based on cumulative capacity as used in traditional experience curves are shown to be vastly superior. Furthermore the effectiveness of increasing weights for more recent data is tested using Weighted Least Squares with exponentially increasing weights. This results in forecasts that are less biased, though have increased spread when compared to Ordinary Least Squares.