Hourly total gaseous mercury (TGM) concentrations at three monitoring sites (receptors) in Ontario were predicted for four selected periods at different seasons in 2002 using the Stochastic Time-Inverted Lagrangian Transport (STILT) model, which transports Lagrangian air parcels backward in time from the receptors to provide linkages to the source region in the upwind area. The STILT model was modified to deal with Hg deposition and high stack Hg emissions. The model-predicted Hg concentrations were compared with observations at three monitoring sites. Estimates of transport errors (uncertainties in simulated concentrations due to errors in wind fields) are also provided that suggest such errors can reach approximately 10% of simulated concentrations. Results from a CMAQ chemical transport model (CTM) simulation in which the same emission and meteorology inputs were used are also reported. The comparisons show that STILT-predicted Hg concentrations usually agree better with observations than CMAQ except for a subset of cases that are subject to biases in the coarsely resolved boundary conditions. STILT captures high frequency concentration variations better than the Eulerian CTM, due to its ability to account for near-field influences that are not resolved by typical grid sizes in Eulerian CTMs. Thus it is particularly valuable for the interpretation of plumes (short-term concentration variations) that require complex sub-grid treatments in Eulerian models. We report quantitative assessments of the relative importance of different upstream sources for the selected episodes, based on emission fluxes and STILT footprints. The STILT simulations indicate that natural sources (which include re-emission from historical anthropogenic activities) contribute much more than current-day anthropogenic emissions to the Hg concentrations observed at the three sites.