[Paper Review] If You Like It, GAN It—Probabilistic Multivariate Times Series Forecast with GAN

Alireza Koochali|arXiv (Cornell University)|Jul 8, 2021

Stock Market Forecasting Methods28 references29 citations

TL;DR

This paper proposes ProbCast, a conditional GAN-based probabilistic multivariate time-series forecasting model that transforms deterministic point forecast models into more accurate probabilistic counterparts. By leveraging a deterministic model's architecture to guide GAN component design, the method achieves stable training and superior performance on electricity and exchange-rate datasets, with CRPS∗ improvements of 1.7% and 17.2% respectively.

ABSTRACT

The contribution of this paper is two-fold. First, we present ProbCast—a novel probabilistic model for multivariate time-series forecasting. We employ a conditional GAN framework to train our model with adversarial training. Second, we propose a framework that lets us transform a deterministic model into a probabilistic one with improved performance. The motivation of the framework is to either transform existing highly accurate point forecast models to their probabilistic counterparts or to train GANs stably by selecting the architecture of GAN’s component carefully and efficiently. We conduct experiments over two publicly available datasets—an electricity consumption dataset and an exchange-rate dataset. The results of the experiments demonstrate the remarkable performance of our model as well as the successful application of our proposed framework.

Motivation & Objective

• To develop a probabilistic multivariate time-series forecasting model that captures uncertainty in predictions.
• To address the instability and high hyperparameter sensitivity of training GANs for time-series forecasting.
• To propose a framework that transforms existing deterministic forecasters into more accurate probabilistic models without compromising performance.
• To demonstrate that GAN-based probabilistic forecasting can outperform deterministic baselines in both small and large multivariate datasets.

Proposed method

• ProbCast employs a conditional GAN framework to model P(Xt+1|Xt, ..., X0), learning the conditional probability distribution of future values.
• The generator uses a noise vector z and a deterministic backbone (GRU + MLP) to produce multivariate forecasts, with architecture derived from a pre-trained deterministic model.
• The discriminator uses a GRU block followed by dense layers to distinguish real future sequences from generated ones, trained via adversarial loss.
• A hybrid loss function combines standard GAN loss with a CRPS∗-based objective to improve calibration and sharpness of probabilistic forecasts.
• The framework uses a genetic algorithm to search for optimal GAN component architectures, reducing the search space by reusing the deterministic model’s structure.
• Hyperparameters such as input window size, noise size, and GRU layer configurations are optimized per dataset using this framework.

Experimental results

Research questions

• RQ1 Can a conditional GAN be effectively trained for multivariate time-series forecasting with stable convergence?
• RQ2 Can a deterministic point forecast model be systematically transformed into a more accurate probabilistic model using GANs?
• RQ3 Does the proposed framework reduce the architectural search space for GANs while improving forecast performance?
• RQ4 How does the resulting probabilistic model compare to its deterministic predecessor in terms of calibration and sharpness on real-world datasets?

Key findings

• ProbCast achieved a CRPS∗ of 232.00 on the electricity dataset, outperforming its deterministic baseline (235.96), representing a 1.7% improvement.
• On the smaller exchange-rate dataset, ProbCast achieved a CRPS∗ of 8.66×10⁻³, compared to 1.04×10⁻² for the deterministic model, a 17.2% relative improvement.
• The framework successfully transformed a deterministic model into a probabilistic one with improved accuracy, demonstrating that architectural priors from deterministic models enhance GAN training stability and performance.
• The method generalizes across datasets with varying feature counts and data sizes, showing robustness on both high-dimensional (electricity) and low-dimensional (exchange-rate) time-series.
• The use of CRPS∗ as a metric enabled direct comparison between deterministic and probabilistic models, with ProbCast showing better calibration and sharpness.
• The genetic algorithm-based hyperparameter search efficiently identified optimal architectures, validating the framework’s ability to reduce the burden of GAN hyperparameter tuning.