Tail Instability and Scale Control in Conditional Normalizing Flows for Treasury Yield-Curve Scenarios
A conditional normalizing flow can produce extreme scenarios even when its point forecasts appear reasonable. We compare two fitted models for daily Treasury yield changes at eleven maturities reconstructed from Federal Reserve curve parameters. Both use an LSTM, an auxiliary location head, and an affine residual flow. On a previously unused 60-date window, the frozen Student-t model generated an absolute standardized coordinate of 34.3 million. A Gaussian-base model with tighter coupling scale control reached 14.72. Each generated 600,000 vectors. The original model's scenario-mean MSE was 33,908.95, while its location-head MSE was 0.3376. The stabilized model's two estimates remained close. Across five development windows, the stabilized maximum was 17.33 among three million vectors. Stability did not imply better prediction: historical simulation had better distributional scores and coverage closer to nominal on the new window. The original flow also had a better log score despite its extreme samples. Frozen-model evaluation is reproducible, but the comparison does not isolate architectural effects or establish a general forecasting advantage.