Certain models are specifically crafted to process time series data, such as ARIMA, ETS, and N-Linear. Conversely, some models struggle with trends, e.g., Decision Tree, Linear Tree, GBM, and also various neural network models.
Instead of predicting absolute values that were not observed in the training data, we can allow the model to forecast relative changes, i.e., $({y_{t+h}} - y_t)/y_t$ rather than $y_{t+h}$. For example, if the training data had values 100 and 200, the increase is 100%. A model that learned to predict a 100% increase can output 400 as the future value, even though such a range of numbers was never observed during training.
During evaluation, we convert the predicted percentage changes back to absolute values and compare them with the actual values. The use of percentage change is determined by the hyper parameter optimization.
For input data, we apply transformations as explained in stationary time series. By utilizing differenced historical values, the model focuses on learning from the changes of featuers than their absolute values. When diff is combined with log-transformation, it effectively captures relative changes in the features since $\log(x_{t+h}) - \log(x_t) = \log(x_{t+h}/x_t)$. However, whether those will be actually used will depend on hyper parameter optimization.