Two-stage Odd Residual Flows for Mean-Preserving Probabilistic Time Series Forecasting
TORF keeps the point forecast’s mean fixed while modeling residual uncertainty with an odd-function normalizing flow.
The paper splits forecasting into two stages: a deterministic model first predicts the mean, then a restricted flow learns the residual distribution around it. The odd residual design is meant to preserve that mean without Monte Carlo sampling. The authors report state-of-the-art NMAE for deterministic accuracy and strong CRPS density-estimation results across short- and long-horizon forecasts. ArXiv · AI/CL/LG's note
The paper splits forecasting into two stages: a deterministic model first predicts the mean, then a restricted flow learns the residual distribution around it. The odd residual design is meant to preserve that mean without Monte Carlo sampling. The authors report state-of-the-art NMAE for deterministic accuracy and strong CRPS density-estimation results across short- and long-horizon forecasts. ArXiv · AI/CL/LG's note
score 4