This paper presents a novel detrending algorithm that allows long-term natural gas demand signals to be used effectively to generate high quality short-term natural gas demand forecasting models. Short data sets in natural gas forecasting inadequately represent the range of consumption patterns necessary for accurate short-term forecasting. In contrast, longer data sets present a wide range of customer characteristics, but their long-term historical trends must be adjusted to resemble recent data before models can be developed. Our approach detrends historical natural gas data using domain knowledge. Forecasting models trained on data detrended using our algorithm are more accurate than models trained using nondetrended data or data detrended by benchmark methods. Forecasting accuracy improves using detrended longer-term signals, while forecast accuracy decreases using non-detrended long-term signals.