TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers

Yilie Huang

In the world of quantitative finance, diffusion models have become a go-to tool for generating realistic market scenarios. But there's a catch: to be useful in large-scale risk simulations, these models need to be fast—often running with just a handful of steps. This "low-budget" sampling introduces numerical errors that can silently corrupt the very metrics—like tail risk or asset correlations—that managers rely on.

A new paper from Yilie Huang at The Hong Kong Polytechnic University introduces TAPES (Terminal Adjoint-Propagated Error Scheduling), a method that rethinks how we choose sampling timesteps to ensure that every "click of the clock" counts toward the final financial report.

TL;DR

TAPES optimizes the sampling grid of diffusion models by mathematically "transporting" local numerical errors to see how they impact final financial reports. By accounting for how these errors reinforce or cancel each other out, it significantly improves the accuracy of low-cost scenario generators without changing the underlying model.

What Problem Does This Paper Address?

Diffusion models generate data by reversing a noise process through a series of small steps. In finance, we use these to simulate future asset returns. Because these simulations often run inside massive Monte Carlo loops, they must be extremely cheap—typically using only 4 to 32 steps (Function Evaluations, or NFE).

At such low budgets, discretization error is a major problem. A "local" error made early in the sampling process might have a huge impact on the final covariance matrix, while a similar-sized error later on might be negligible. Standard methods for choosing timesteps often treat all errors equally or focus on general "path quality." TAPES addresses the gap between general numerical accuracy and the specific terminal reports (like Value-at-Risk, drawdown, or volatility persistence) that financial pipelines actually consume.

Methods

TAPES is an offline grid construction technique. It doesn't retrain the model; it just finds the best "schedule" for the steps. The process works in three main stages:

  1. Richardson Comparison: It estimates the "signed local defect" (the error) at each step by comparing one full step to two half-steps.
  2. Adjoint Propagation: It uses a "terminal-report adjoint" to transport these local errors to the end of the simulation. This tells us exactly how a tiny error at step 5 affects, for example, the final portfolio variance.
  3. Optimization: It minimizes the squared norm of the sum of these transported errors. This is a crucial distinction from standard methods—by summing before squaring, TAPES accounts for the fact that errors across different steps can either reinforce or cancel each other out.

Finally, it performs a short "direct refinement" on a small development dataset to polish the grid before it is frozen for deployment.

Datasets

The authors evaluated TAPES on two distinct regimes:

  • Regime-SV: A synthetic five-asset simulation featuring regime-switching stochastic volatility (moving between "calm" and "stress" states).
  • Mag7: Real-world historical data for the "Magnificent Seven" tech stocks (AAPL, MSFT, NVDA, AMZN, GOOGL, META, TSLA), using data from 2012 through mid-2026 for evaluation.

Key Insights and Ideas

  • The Reinforcement Regime: One of the paper's most striking findings is that numerical errors in diffusion models tend to reinforce each other. Traditional "diagonal" error metrics, which just sum up the squares of local errors, ignore these cross-step interactions. TAPES explicitly targets this reinforcement.
Figure 2: Final-grid mechanism at 𝐾= 8. (a) Deployed incre- ments Δ𝑘/𝑇for Mag7; UNIF dotted. (b, c) Evaluation-panel means and pointwise 95% intervals for the diagonal term 𝐷, signed cross term 𝐶, and 𝑆= 𝐷+ 𝐶in normalized design- report coordinates. 𝐶> 0 indicates cross-step reinforcement.
  • Goal-Oriented Scheduling: The "best" grid depends on what you are measuring. A grid optimized for image quality isn't necessarily the best for calculating a 5% tail quantile. TAPES allows the grid to be "report-aware."
  • First-Order Theory: The authors provide a theoretical proof showing that their transported signed sum is the dominant (first-order) term in the final report error, providing a solid mathematical foundation for the algorithm.

Key Results

TAPES consistently outperformed state-of-the-art baselines like ART (Adaptive Reparameterized Time) and AYS (Align Your Steps) across all tested budgets.

  • Error Reduction: On the Regime-SV dataset, TAPES reduced audit-report errors by 10–11% on average. On the more challenging Mag7 dataset, it still provided a consistent ~1% improvement.
  • Consistency: TAPES achieved the lowest error in all 24 dataset-budget-report combinations tested.
Figure 1: Paired audit-error reductions across deployment budgets; positive values favor TAPES over the ART and AYS baselines. Shading gives pointwise 95% paired-bootstrap in- tervals on the Table 1 panel (iid for Regime-SV; non-circular blocks of four Mag7 contexts). All 24 audit intervals exclude zero; because budgets reuse items, the intervals are not si- multaneous.
  • Financial Impact: In an exploratory test, the authors used these grids to drive a portfolio allocator. The TAPES-optimized scenarios led to better Sharpe and Sortino ratios, suggesting that better numerical fidelity translates directly into better financial decision-making.

Limitations and Future Work

While powerful, TAPES has some boundaries:

  • Offline Cost: The grid optimization is an "offline" cost. While it only takes a few minutes on a laptop, it must be done once for each model and budget.
  • Frozen Generator: TAPES improves the sampling of a model, but it cannot fix a fundamentally poorly trained "score network." It ensures fidelity to the generator, not necessarily to the "true" market law.
  • Scope: The current work focuses on Euler and Heun solvers. Extending this to more complex, higher-order solvers remains a direction for future research.

In conclusion, TAPES demonstrates that in the high-stakes world of financial simulation, how you spend your "numerical budget" is just as important as how much budget you have. By looking ahead to the final report, TAPES makes low-NFE diffusion models significantly more reliable for risk management.

Summary updated

Abstract

Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines, where numerical time-grid error can distort terminal quantities such as covariances, tails, and drawdown proxies. We study offline grid construction for a frozen probability-flow ODE: the score network, conditioning, Euler solver, report normalization, and deployment NFE remain fixed. TAPES (Terminal Adjoint-Propagated Error Scheduling) estimates signed local defects by Richardson comparisons, transports them through the remaining discrete flow and a terminal-report adjoint, and optimizes the squared norm of their sum before a capped direct refinement. We show that this transported signed sum is the first-order terminal-report error and quantify, under an aligned-defect model, the inefficiency of diagonal criteria. Across six budgets on a regime-switching simulation and chronologically held-out Mag7 contexts, TAPES gives the lowest design- and audit-report errors in all 24 dataset--budget--report rows within the reported five-pipeline panel. Cross-budget mean relative audit-error reductions are about 10--11% on Regime-SV and about 1% on Mag7, with Bonferroni intervals excluding zero; all displayed winners persist under a doubled-resolution reference, and a small Regime-SV Heun probe preserves the TAPES-first audit ordering. TAPES changes only the offline schedule and retains the prescribed deployment NFE; claims concern numerical fidelity to frozen generators.
BibTeX
@misc{huang2026tapes,
  author = {Yilie Huang},
  title = {{TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers}},
  year = {2026},
  eprint = {2611.16484},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16484},
}
Figure