Probabilistic Neural Forecasts for Energy Markets: A Comparison with Econometric Baselines and Forward Curves

Abstract Hydrocarbon supply & trading daily price calls drive planning, hedging, and refinery runs, yet crude markets can swing on a headline and make forecasts fragile. We test whether a multiscale probabilistic neural forecaster can improve daily Brent calls over common industrial forecasting baselines on real data. Using Dated Brent as the case, we build a robust calendar-faithful univariate workflow and compare Automated Neural Hierarchical Interpolation for Time Series (AutoNHITS) against a no-change naïve forecast, Autoregressive Integrated Moving Average (ARIMA), and leading Price Reporting Agency (PRA) published forward curves. Models are trained and selected on time-ordered splits with no resampling and no exogenous inputs. AutoNHITS is trained under a Normal-likelihood objective, its point forecast is the median of the learned predictive distribution, and it also provides 75% and 90% zero-floored prediction bands from the same fit. Across 240 out-of-sample days in 2025, AutoNHITS posts the highest mean and median accuracy (94.52% and 95.10%), the top win-rate, and safer tails with lower 95th/99th Absolute Percentage Error (APE) than the PRA forward curve. Gains hold across volatility regimes and grow on larger daily moves. The outcome is a practical, drop-in forecasting approach that preserves the native calendar, requires no exogenous drivers, and pairs cleanly with forward-curve practice to support planning, hedging, and risk decisions.