问题
I am working with Deepar and trying to get a better understanding of the quantile values returned. From the documentation, the likelihood hyperparameter explains that: ...provide quantiles of the distribution and return samples.
If I look at a single data point the quantiles returned are linear. E.g. the 0.1 quantile has the lowest predicted value and 0.9 quantile has the highest predicted value. I am having trouble understanding this. If these are samples from the distribution, shouldn't they look similar to the distribution selected with the likelihood hyperparameter (negative-binomial in my case)?
回答1:
DeepAR returns probabilistic forecasts in terms of quantiles: by default, the 0.1, 0.2, 0.3, ..., 0.9 quantiles are returned. This means that, according to the model, in each future time step you have 10% chance of observing something lower than the 0.1 quantile, 20% chance of observing something lower than the 0.2 quantile, and so on. Quantiles are in fact in order, and they must be by definition of quantile. Hope this clarifies is a bit!
来源：https://stackoverflow.com/questions/58325923/deepar-prediction-quantiles-explained