Evidential Deep Learning

Most neural networks predicting, say, tomorrow's return output a single number and stop there — no sense of how confident that number is. The usual fix, training an ensemble of many networks and looking at how much they disagree, is expensive: you need dozens of full training runs. Evidential deep learning gets an uncertainty estimate out of a single network trained once, by changing what the network outputs.

Instead of predicting a point value, the network predicts the parameters of a distribution over what the true value could be — for regression, typically the parameters of a Normal-Inverse-Gamma distribution, which itself describes a distribution over "what mean and variance would explain this data." Think of it as the network saying not just "I predict 2%" but "here's how much evidence I have for that prediction and how spread out the plausible alternatives are," all computed in one pass rather than by training many models and comparing them.

This separates two kinds of uncertainty in one shot: how spread out the predicted distribution is (noise inherent in the data — aleatoric) versus how little evidence backs the prediction (model uncertainty — epistemic, which should shrink with more relevant training data and spike on unfamiliar inputs, like a stock in a sector the model rarely saw). The training loss is designed so overconfident wrong predictions are penalized, encouraging the reported uncertainty to actually track reliability rather than being an afterthought bolted onto point predictions.