Major cyclones are still occasionally poorly predicted with present-day operational weather prediction models in regions downstream of sparse observational assets. One such cyclone made landfall on the Oregon coast on February 7th, 2002, and produced widespread damage. Operational short-term forecasts of this cyclone contained both position and intensity errors. Using the PSU/NCAR MM5 mesoscale model and its adjoint, it is shown that large short term sensitivity gradients of forecast variables with respect to initial conditions existed. This supports the idea that small errors in initial conditions likely led to large forecast deviations, and that initial condition error was likely the cause of the forecast failure. A four dimensional variational data assimilation (4DVAR) system using the MM5 was employed to further examine the forecast failure of one operational MM5 ensemble member. Various optimal initial conidtions were constructed using different cost functions, and these optimal condtions were compared to eachother and the original forward model run. It appears that the perturbations made to the original initial conidtions are relatively small, but grow rapidly throughout the forecast. Furthermore, these various initial conditions and correpsonding forecasts were compared to a posteriori observational analysis in order to determine which more closely matched reality. Although the forecasts are forced to be accurate through the cost function minimizaiton procedure within the 4DVAR system, actual forecast trajectories originating from optimal initial conditions aren't necessarily more realistic. This similarity to reality seems to depend on the choice and extent of the cost function. Sensitivity gradients of various operational MM5 ensemble members will also be compared to determine their nature and similarities to eachother. Finally, comments on the feasibility of these methods for operational, real-time use will be made.