Paper: arXiv 2305.16274

Authors: Zacharia Issa, Blanka Horvath, Maud Lemercier, Cristopher Salvi

Abstract

Neural SDEs are continuous-time generative models for sequential data. State-of-the-art performance for irregular time series generation has been previously obtained by training these models adversarially as GANs. However, as typical for GAN architectures, training is notoriously unstable, often suffers from mode collapse, and requires specialised techniques such as weight clipping and gradient penalty to mitigate these issues. In this paper, we introduce a novel class of scoring rules on pathspace based on signature kernels and use them as objective for training Neural SDEs non-adversarially. By showing strict properness of such kernel scores and consistency of the corresponding estimators, we provide existence and uniqueness guarantees for the minimiser. With this formulation, evaluating the generator-discriminator pair amounts to solving a system of linear path-dependent PDEs which allows for memory-efficient adjoint-based backpropagation. Moreover, because the proposed kernel scores are well-defined for paths with values in infinite dimensional spaces of functions, our framework can be easily extended to generate spatiotemporal data. Our procedure permits conditioning on a rich variety of market conditions and significantly outperforms alternative ways of training Neural SDEs on a variety of tasks including the simulation of rough volatility models, the conditional probabilistic forecasts of real-world forex pairs where the conditioning variable is an observed past trajectory, and the mesh-free generation of limit order book dynamics.

Complexity vs Empirical Score

  • Math Complexity: 9.0/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper is highly theoretical, featuring advanced mathematics like signature kernels, infinite-dimensional pathspace analysis, and strict properness proofs, but it lacks any empirical implementation details such as code, backtests, or datasets, focusing instead on theoretical guarantees and high-level descriptions of applications.

Research Flowchart

  flowchart TD
  A["Research Goal: Non-adversarial<br>Training of Neural SDEs"] --> B{"Introduce Novel Scoring Rule<br>based on Signature Kernels"}
  B --> C["Derive Theoretical Guarantees<br>Strict Properness & Consistency"]
  C --> D["Formulate Generator-Discriminator<br>as Linear Path-Dependent PDEs"]
  D --> E["Apply Memory-Efficient<br>Adjoint Backpropagation"]
  E --> F["Key Outcomes"]
  
  subgraph F
      F1["Stable Training<br>No Mode Collapse"]
      F2["Spill-over to<br>Spatiotemporal Data"]
      F3["State-of-the-Art Performance<br>on Rough Volatility & LOB"]
  end

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a non‑adversarial training method for Neural SDEs using signature kernel scores as the objective.

Source sentence
we introduce a novel class of scoring rules on pathspace based on signature kernels and use them as objective for training Neural SDEs non-adversarially.

Assumes 4

  • Technical assumption needed to ensure characteristicness of the signature kernel. model
    Source sentence
    This is a technical assumption needed to ensure characteristicness of the signature kernel. See Proposition (3.1).
  • Paths x, y are almost‑everywhere differentiable so that the signature kernel satisfies a linear hyperbolic PDE. model
    Source sentence
    which reduces to a linear hyperbolic PDE in the when the paths x, y are almost-everywhere dif‑ferentiable.
  • The space X is equipped with a topology that makes the signature continuous. model
    Source sentence
    From now on we endow X with a topology with the respect to which the signature is continuous; see [CT22] for various choices of such topologies.
  • A target random variable Xtrue with law PXtrue is available for training. data
    Source sentence
    We are given a target X-valued random variable Xtrue with law PXtrue.

Data

  • Geometric Brownian motion synthetic dataset synthetic unknown
    Source sentence
    We chose µ = 0, σ = 0.2 and generated time-augmented paths of length 64 over the grid ∆= {0, 1, . . . , 63} with dt = 0.01.
  • Foreign exchange (FX) currency pairs dataset market-prices unknown
    Source sentence
    the third is a multidimensional example with foreign exchange (forex, or FX) currency pairs, which was chosen not only because of the relevance and capitalisation of FX markets but also due to its well-known intricate complexity.
  • NASDAQ limit order book data · NASDAQ public exchange [NMK+18] order-book public
    Source sentence
    We fit our model on real LOB data from the NASDAQ public exchange [NMK+18] which consists of about 4M timestamped events with L = 10 price levels.

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • The generator trained with the full signature kernel score (ϕsig) achieves the lowest KS scores and Type I error at the 5 % significance level on the geometric Brownian motion benchmark. performance
    Source sentence
    Here the generator trained with ϕsig performs the best, achieving a Type I error at the assumed confidence level.
  • The proposed non‑adversarial training significantly outperforms alternative training methods (including SDE‑GAN) across multiple tasks such as rough volatility simulation, FX forecasting, and limit order book generation. performance
    Source sentence
    significantly outperforms alternative ways of training Neural SDEs on a variety of tasks including the simulation of rough volatility models, the conditional probabilistic forecasts of real‑world forex pairs
  • Average KS scores for each spatiotemporal marginal of the NASDAQ limit order book are reported, showing competitive performance of the Neural SPDE model trained with signature kernel scores. performance
    Source sentence
    Figure 3: KS test average scores for each spatiotemporal marginal, 100 runs, NASDAQ data.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 15 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 6 papers in the archive (5 by independent authors): Universal randomised signatures for generative time series … (uses its method or data)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.