Portfolio policies are trained on predictions of future returns, but these predictions are uncertain. A common intuition is that supplying a policy with both the predicted mean and a calibrated uncertainty summary should improve risk-sensitive portfolio choice. We test this intuition using a controlled experimental design that separates more inputs from more information . We represent forecasts as a discretized time–asset forecast field and compare (i) rolling-window and Kalman-filter beliefs with explicit calibration, and (ii) multiple policy families spanning neural operators and strong non-operator baselines: DeepONet, a modern factorized Fourier neural operator (F-FNO), a variational operator baseline (VINO-style), a Transformer encoder, and a residual MLP. For each architecture we run matched-dimension ablations in which the uncertainty channel is replaced by variance-matched random noise. Across datasets and evaluation protocols (time-split cross-validation and held-out crash/OOD stress tests), calibrated uncertainty features do not reliably improve out-of-sample certainty equivalent (CE) relative to mean-only inputs, and performance is often indistinguishable from the noise-control. Mechanistic diagnostics (input gradients, representation sensitivity, and attribution) indicate that uncertainty channels frequently become functionally inactive under the training objective and data regime. Our conclusions are empirical and limited to the tested architectures and settings; we provide diagnostics and reporting templates to make uncertainty-usage failures observable and reproducible.
WonChan Cho (Sun,) studied this question.