Dead fine fuel moisture content (FM) is one of the most important determinants of fire behavior. Fire scientists have attempted to effectively estimate FM for nearly a century, but we are still lacking broad scale evaluations of the different approaches for prediction. Here we tackle this problem by taking advantage or a recently compiled global fire behavior database (BONFIRE) gathering 1603 records of 1h (i.e., <6 mm diameter or thickness) dead fuel moisture content from measurements before experimental fires. We compared the results of models routinely used by different agencies worldwide, empirical models, semi-mechanistic models and also non-linear and machine learning approaches based on either temperature and relative humidity or vapor pressure deficit (VPD). A semi-mechanistic model based on VPD showed the best performance across all FM ranges and a historical model developed in Australia (MK5) was additionally recommended for low fuel moisture estimations. We also observed significant differences in FM dynamics between vegetation types with FM in grasslands more responsive to changes in atmospheric dryness than woody ecosystems. The addition of computational complexity through machine learning is not recommended since the gain in model fit is small relative to the increase in complexity. Future research efforts should concentrate on predictions at low FM (<10 %) as this is the range most significant for fire behavior and where the poorest model performance was observed. Model predictions are available from https://hcfm.shinyapps.io/shinyfmd/.