Naturally Fractured Reservoirs (NFR) hold a significant fraction of remaining petroleum reserves. Recovery factors from NFR are usually less than in conventional reservoirs due to associated high uncertainty throughout the characterisation and modelling phases. This particularly includes the modelling and upscaling of the fracture domain using Discrete Fracture Networks (DFN). Computer assisted history matching and prediction is becoming increasingly popular as they help finding multiple history-matched models and probabilistic forecasts. Therefore, the associated uncertainty can now be quantified in a limited time frame. However, the results of a history match are known to depend on initial reservoir properties, including fracture permeability and matrix shape factors. Geological uncertainty in these two factors is exacerbated by the DFN upscaling errors. We show how DFN modelling can be used to increase geological prior knowledge and hence produce more geologically consistent models. To highlight DFN upscaling errors, we use a realistic dataset from an onshore fractured reservoir to show how the DFN upscaling error could propagate through to the history matching phase. We compare history matching of three models with different DFN upscaling processes. Results from state-of-the-art assisted history matching and prediction were found to depend on the static properties and particularly the computation of effective fracture permeability during DFN upscaling. This upscaling error alone leads to very different reservoir models, despite the best history matched models being of comparable quality. Hence, this leads to more uncertainty in reservoir production forecast. The identification of DFN upscaling errors is therefore crucial for better uncertainty quantification in reservoir simulation of NFR.