The empirical relationship between crude oil prices and petroleum inventories has been exploited in a number of short-term oil price forecasting models. Some of the models are based on the perception that an unexpected inventory level indicates an imminent price change, implicitly assuming that imbalances between crude oil supply and demand affect inventories before price. The last years these "relative inventory level" models have failed, in the sense that they have consistently under-predicted the oil price. Some relate the failure to an apparent inversion in the relationship between price and inventories. Ye, Zyren et al. (2006) find that the "unusual positive relationship" may indicate that the oil market is in a state of transition. Merino and Ortiz (2005) find that futures market activity may explain parts of the forecast failure. Tchilinguirian (2006) observes that the price-inventory relationship has moved "out of line", and suggests that the reason may be a shift in the storage supply curve.This thesis asks if there is a break in the relative inventory level model that can be attributed to the simplistic representation of the price-inventory relationship, or if the break is rather in the oil price process, unrelated to the price-inventory relationship.The main find is a slow shift in the oil price process between 2002 and 2007, towards a new regime with a higher mean, higher variance and lower persistence. The investigated relative inventory models fail because they assume that the oil price reverts to a constant mean. Furthermore, it is found that whilst the price-inventory relationship is not broken, the statistical properties of the linear relative inventory model may be considerably improved by a quadratic log-log respecification. The findings have some consequences. Firstly, some of the lost forecasting power may be restored by introducing a model component that represents a slow shift in the long-run oil price. Although this fix does not explain the causes of the shift, it absorbs some of the distortions from the shift, and enables estimation of the parameters by the use of simple regression methods. Secondly, the proposed quadratic log-log specification allows for a more proper statistical treatment of uncertainty, and is well-specified enough to be used for hypothesis testing and computation of confidence bands. Finally, the identified regime shift resembles the shifts that Videgaray-Caso (1998) found to come about every 11th year or so, before reverting back after another 4 or 5 years.A simple approach is followed to answer the research questions: The point of departure is a dissection of the most parsimonious amongst the reportedly successful relative inventory models. An explicit formulation of the embedded economic model justifies a slightly modified economic model that is more aligned with theory, and presumably more robust in the face of data. The two economic models lead to two econometric models: The baseline model, basically equal to the dissected model, and the competing quadratic log-log model. Both econometric models are then estimated using ordinary least squares and evaluated on the original dataset (1992-2003) and a more complete dataset (1992-2007), in order to investigate if the simple respecification remedies any break. Next, a version of the proposed model that allows for a smooth shift in the intercept and autoregressive parameters is specified and estimated using maximum likelihood and nonlinear least squares on the complete dataset, primarily in order to investigate the hypothesis that there is a break in the oil price process.