Paper: arXiv 2610.14584

Authors: Arin Mohanty

Abstract

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.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel approach to stabilizing conditional normalizing flows for yield curve scenarios, backed by rigorous empirical testing on real-world Treasury data. While the mathematical details of the flow architecture are present, the focus is on their application and empirical performance. The comparison with benchmarks and detailed metrics contribute to its strong empirical rigor.

Research Flowchart

  flowchart TD
    A[Research Goal: Control Tail Instability in Conditional Normalizing Flows for Yield-Curve Scenarios] --> B{Methodology: Compare Two Models};
    B --> C{Model 1: Frozen Student-t with Affine Residual Flow & LSTM};
    B --> D{Model 2: Gaussian-Base with Tighter Coupling Scale Control, Affine Residual Flow & LSTM};
    C & D --> E[Data Input: Daily Treasury Yield Changes (11 maturities) from Fed Curve Parameters];
    E --> F[Computational Process: Model Fitting & Scenario Generation (600k vectors/model, 60-day window)];
    F --> G{Key Findings: Extreme Scenario Control & Prediction Performance};
    G --> H[Outcome 1: Frozen Student-t: Max Std. Coord 34.3M, Scenario-mean MSE 33908.95];
    G --> I[Outcome 2: Gaussian-Base: Max Std. Coord 14.72, Stabilized estimates close];
    G --> J[Outcome 3: Stability did not imply better prediction: Historical Simulation better distributional scores];
    J --> K[Outcome 4: Original Flow had better log score despite extreme samples];