Research map / Time series / Causal and generative time series

Generating Financial Time Series by Matching Random Convolutional Features

NeurIPS 2026 · Time series › Causal and generative time series · approach: State-space and latent dynamics forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains conditional financial time series generators by matching the means of a new differentiable random convolutional feature map (SOCK, a soft version of Hydra/Rocket-style kernel competitions) resampled during training, outperforming signature feature matching and Diffusion-TS on small single-path datasets.

Problem

Generating realistic future financial time series conditional on recent history when only a single short historical path is available, where adversarial training overfits and existing signature-based feature matching may miss relevant properties.

Approach

Replace signature features with random convolutional features made differentiable: SOCK applies Hydra-style random kernel competitions over grouped (width-2) convolutions, replaces hard argmax with softmax win probabilities, and pools them via soft-deviation; the feature map is resampled during generator training so the generator matches feature means across many random maps.

Key results

• SOCK attains the best average rank (1.49) over all dataset-metric pairs for generative modelling, followed by rSig (2.80).
• SOCK matches MultiRocket on 112 UCR datasets while using an order of magnitude fewer random features and a single pooling, and outperforms Hydra.
• Resampling the SOCK feature map during training improves SRNN from 0.126 to 0.117 on TGH; for rSig it improves SRNN from 0.416 to 0.296.

Surprising findings

• SOCK already trains good generators without feature-map resampling, while rSig gains greatly from it.
• Gains on real daily-return datasets are smaller than on synthetic data, attributed to weak autocorrelation and noisy evaluation.

Limitations

• Experiments focus on low-dimensional datasets with a single historical path; larger datasets not studied.
• Deliberately simple GRU generator used for all feature-matching methods; stronger architectures untested.
• Ablations run only on the TGH dataset, so results may differ elsewhere.

Experimental setup

Datasets: V1 V10 TGH SV FGN FI PH ST IT IDX FX CRY

Benchmarks: SV V10 TGH FGN FI PH ST IT IDX FX CRY UCR 112 datasets

Metrics: SIG MLP RNN SRNN ACF CCF CVM ES ED PES

Baselines: Sig rSig cSig (Sig-Wasserstein GAN) Diffusion-TS Hydra MultiRocket Rocket truncated RBF signature kernel GAK

Builds on: Hydra Rocket MultiRocket signature feature matching (Sig-Wasserstein GAN) randomized signatures Diffusion-TS

Closest papers by meaning

• Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.87)
• Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.87)
• Time Weaver: A Conditional Time Series Generation Model (ICML 2024 · similarity 0.86)
• A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking (NeurIPS 2025 · similarity 0.85)
• Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series (NeurIPS 2024 · similarity 0.85)
• Time Series Representations with Hard-Coded Invariances (ICML 2025 · similarity 0.85)