Icing on wind turbine blades causes significant production losses for wind energy in cold climate. Next-day forecasts of these production losses are crucial for the power balance in the electrical grid and for the trading process, but they are uncertain due to lack of understanding of, and simplifications, in the modelling chain. In the present work, uncertainties in the modelling chain for icing related production losses are addressed with the aim to increase the utility of next-day production loss forecasts. Probabilistic and machine learning methods are applied both to improve the forecast skill and to estimate reliable forecast uncertainties. The different methods enable uncertainties in different parts of the chain to be addressed. A Numerical Weather Prediction (NWP) ensemble captures uncertainties in the initial conditions of the forecasts while a neighbourhood method describes uncertainties in the spatial representation of the NWP forecast at the exact locations of the wind parks. An icing model ensemble is generated in order to address uncertainties in the icing model parameters. Finally, machine learning approaches are employed to both deterministically and probabilistically address uncertainties in the modelling chain. Production data from wind parks in Sweden were used to evaluate all methods. The physically based probabilistic methods; the NWP ensemble, the neighbourhood method and the icing model ensemble, increase the forecast skill and provide valuable uncertainty estimations. The largest forecast improvement is obtained when the different probabilistic approaches are combined. On the other hand, machine learning approaches for icing related production losses demonstrate large potential. The probabilistic machine learning method employed generally outperforms every other single probabilistic method mentioned above. By applying the different methods of uncertainty quantification, the utility of icing related production loss forecast in the trading process is improved since related costs can be reduced and usage of the produced power can be optimised. These methods can also be beneficial when planning for site maintenance and for the use of de-icing systems, since icing on the wind turbines are directly or indirectly forecasted. Thus, the improved representations of uncertainties in the modelling chain contributes to an enhanced usage of wind power in cold climates.