Forecasting fatalities from state-based armed conflict is an expanding area of research aimed at supporting early warning systems and informed policy responses. Accurate forecasts are essential to mitigate the severe human and societal consequences of violent conflict. This thesis explores how the performance of a Markov structure model can be enhanced through improved state definitions and methodological refinements. The model captures transitions between varying levels of intensity of the conflict and generates probabilistic forecasts of monthly fatalities. In addition to the baseline random forest approach, the thesis evaluates an alternative learning algorithm (XGBoost), three feature selection strategies (using principal components, all features, or recursive feature elimination), and a random search procedure to optimize state definitions while preserving domain knowledge. The models are trained on data from the Uppsala Conflict Data Program, covering monthly fatalities in 191 countries. The results suggest that in this context, models using random forest, principal components, and optimized states generally perform better than other configurations. Random forest-based models seem to adapt well to various contexts, while principal components presumably get rid of noise in the conflict data. State optimization often yields performance improvements, but predefined states based on expert knowledge still perform competitively.