Sparse Latent Factor Forecaster (SLFF) with Iterative Inference for Transparent Multi-Horizon Commodity Futures Prediction

Commodity futures are volatile. Forecasting across horizons with interpretable drivers remains challenging. We propose the Sparse Latent Factor Forecaster with Iterative Inference (SLFF), a structured prediction latent variable model that combines sparse coding, unrolled optimization, and amortized inference. SLFF explicitly optimizes a sparse latent code to explain multi-horizon futures trajectories and trains an encoder whose outputs are validated against the optimization-based solution before deployment. The method is paired with an information set aware pipeline (vintage macro releases, lag aware fills, leakage checks) and evaluated under rolling origin folds against representative statistical and neural baselines. We provide quantitative criteria for factor labeling and directional diagnostics that account for no change regimes. On Copper and WTI futures (2005-2023), SLFF achieves competitive RMSE and MAE, improves directional skill beyond persistence, and yields factors that are stable across seeds and linked to measurable fundamentals. Code, diagnostics, and information set specifications are released for reproducibility.