TimeMCL is a method for predicting multiple possible future time series by leveraging the Multiple Choice Learning (MCL) paradigm. It uses multiple heads in a neural network and leverages the Winner-Takes-All (WTA) loss to increase prediction diversity. MCL has recently attracted attention for its simplicity and ability to handle uncertain and ambiguous tasks. This paper applies this framework to time series forecasting and presents an efficient method for predicting multiple futures, linking it to an implicit quantization objective. We provide insights into the approach using synthetic data, and evaluate it on real time series data, demonstrating promising performance at a low computational cost.
