Reducing Emissions from Deforestation and forest Degradation (REDD) projects are being designed and implemented across tropical countries, intending to curb the contribution of deforestation to greenhouse gas emissions. An important aspect of REDD implementation is the baseline against which reductions are measured. The baseline estimates the business-as-usual emissions from deforestation and forest degradation. We solve a dynamic model of land conversion from forest to agriculture in the presence of REDD, and assess the performance of four baselines. We show that none of the analysed baselines dominates in all performance aspects, and that the final baseline choice needs to maximise the trade-off between the effectiveness to reduce deforestation, cost-efficiency, and changes in income. The frequently used historical average baseline could be improved by using a forward-looking one, which is shown to better account for the opportunity costs faced by landowners. This result hinges on the ability of the baseline to predict deforestation rates without significant underestimations. We advocate the switch from a single-threshold baseline to a corridor methodology, which would provide continued incentives to reduce deforestation, even during periods of high opportunity costs. We finally show how the selection of certain baseline attributes, such as corridor bandwidth and symmetry, can enhance performance. Reducing Emissions from Deforestation and forest Degradation (REDD) projects are being designed and implemented across tropical countries, intending to curb the contribution of deforestation to greenhouse gas emissions. An important aspect of REDD implementation is the baseline against which reductions are measured. The baseline estimates the business-as-usual emissions from deforestation and forest degradation. We solve a dynamic model of land conversion from forest to agriculture in the presence of REDD, and assess the performance of four baselines. We show that none of the analysed baselines dominates in all performance aspects, and that the final baseline choice needs to maximise the trade-off between the effectiveness to reduce deforestation, cost-efficiency, and changes in income. The frequently used historical average baseline could be improved by using a forward-looking one, which is shown to better account for the opportunity costs faced by landowners. This result hinges on the ability of the baseline to predict deforestation rates without significant underestimations. We advocate the switch from a single-threshold baseline to a corridor methodology, which would provide continued incentives to reduce deforestation, even during periods of high opportunity costs. We finally show how the selection of certain baseline attributes, such as corridor bandwidth and symmetry, can enhance performance.