This study advances the research on the convenience yield of natural gas. Econometric models confirm that air temperature is an important explanatory variable in addition to storage levels. Furthermore, an extended linear model shows that one has to account for a changing cost of physical storage in the spirit of Brennan (1958). Besides this, an alternative regime-switching model for the convenience yield helps to put in perspective a prominent finding by Fama, and French (1987). That is, given binding capacity constraints for gas storage, the variance of the futures' basis will increase rather than decrease with the storage levels. Finally and most importantly, robustness tests demonstrate that the extended linear model produces the most viable forecasts and that these forecasts can help to amend the performance of reduced-form models for the gas spot price.transparency increase are now on the agenda. One fact underpinning the rising interest in standardized contracts is the recent start of gas trading at two main energy exchanges in this region. In July 2007, the German EEX (European Energy Exchange) launched its gas trading platform and less than a year later, in May 2008, the Scandinavian Nord Pool followed suit. Further support can be found by looking at the traded volumes of gas contracts on the emerging continental marketplaces: APX Group (Amsterdam Power Exchange), the most liquid gas trading place on the mainland, reported a volume increase of 74% in gas contracts for delivery at the Dutch gas hub TTF and a 16-fold increase in volumes for the Belgian hub Zeebrugge. 1 Along with the importance of exchange-traded gas grows the necessity to find accurate pricing models for the different contracts, and this regularly requires a future spot price. One notable difference between commodities and stocks or bonds is that the current spot price is not directly observable in the market. During the last few decades, a number of general commodity pricing models have been developed and tested. However, the application to natural gas pricing was predominantly tested using reduced-form models such as the well-known two-factor models in Gibson and Schwartz (1990), Schwartz (1997), and Schwartz and Smith (2000). As opposed to equilibrium models, these models build on an exogenously specified stochastic process rather than demand and supply conditions founded in microeconomic theory. They are easily tractable and also capture typical phenomena of the forward curve (e.g. the Samuelson effect 2 ), but they do not offer fundamental economic explanations for the predicted prices.The lack of insight into the real price drivers is a downside of pure reduced-form models because the risk of misspecification is high and the outof-sample performance can be poor. With respect to the Schwartz and Smith model, Carlson, Khokher, and Titman (2007) show that given frictional production adjustments in the oil and gas market, the model will systematically overestimate the prices for oil and gas options. Similarly, Ribeiro and Hodges (200...