A.V. Proletarsky1, D.V. Berezkin2, S.T. Tsaplin3
1–3 Bauman Moscow State Technical University (Moscow, Russia)

1 pav@bmstu.ru, 2 berezkind@bmstu.ru, 3 tsaplin@bmstu.ru

Forecasting extremely short time series (fewer than 10 observations) poses a fundamental challenge in modern data analytics, particularly critical for socio-economic and educational systems with limited historical data. In a rapidly changing environment, including college admissions planning, fluctuations in consumer demand, and monitoring emerging technological trends, traditional statistical forecasting methods (ARIMA, exponential smoothing) show a significant decline in effectiveness due to insufficient data for building reliable statistical models. Existing deep learning architectures are also prone to overfitting in situations with a scarcity of training examples, creating a need for new approaches to short-term forecasting.

The objective of the article is a comprehensive assessment of the applicability of generative adversarial networks (GANs) for addressing the challenges of forecasting extremely short time series, including a comparative analysis of the effectiveness of VanillaGAN, Wasserstein GAN (WGAN), and WGAN with gradient penalty (WGAN-GP) architectures, identifying their limitations, and determining promising directions for practical application in data-scarce conditions.

Experimental findings indicate that all investigated GAN architectures exhibit systemic limitations when working with extremely short time series: the generated sequences are characterized by reduced fluctuation amplitudes (a decrease of 40–60%), smoothed trend components, and insufficient variability compared to the original data. WGAN-GP demonstrated relatively better training stability, yet the quality of generation remains unsatisfactory for direct forecasting. The potential for using generative models as data augmentation tools to create additional training examples has been identified, which could enhance the effectiveness of other forecasting algorithms.

The research results have immediate applications for: improving planning systems in educational institutions where forecasting based on limited admission statistics is required; developing hybrid forecasting systems that combine generative models with classical statistical methods; creating specialized solutions for short-term planning tasks in high-uncertainty environments. The proposed approach to using GANs for data augmentation can be adapted for a wide range of socio-economic forecasting tasks, including market trend analysis, resource planning, and risk assessment under conditions of limited historical information.

Proletarsky A.V., Berezkin D.V., Tsaplin S.T. Application of generative neural networks for predicting extremely short time series. Neurocomputers. 2026. V. 28. № 1. P. 5–16. DOI: https://doi.org/10.18127/j19998554-202601-01 (in Russian)

1. Torres J.F., Hadjout D., Sebaa A. et al. Deep learning for time series forecasting: A survey. Big Data. 2021. V. 9. № 1. P. 3–21.
2. Box G.E.P., Jenkins G.M., Reinsel G.C. et al. Time series analysis: Forecasting and control. Journal of Time Series Analysis. 2015. V. 36. № 4. P. 709–711.
3. Hyndman R.J., Khandakar Y. Automatic time series forecasting: the forecast package for R. Journal of Statistical Software. 2008. V. 27. № 3. P. 1–22.
4. Wen Q., Zhou T., Zhang C. et al. Transformers in time series: A survey. arXiv preprint. 2022. arXiv:2202.07125.
5. Giannone D., Lenza M., Primiceri G.E. Prior selection for vector autoregressions. Review of Economics and Statistics. 2015. V. 97. № 2. P. 436–451.
6. Lim B., Zohren S. Time-series forecasting with deep learning: A survey. Philosophical Transactions of the Royal Society A. 2021. V. 379. № 2194. P. 20200209.
7. Lara-Benítez P., Carranza-García M., Riquelme J.C. An experimental review on deep learning architectures for time series forecasting. International Journal of Neural Systems. 2021. V. 31. № 3. P. 2130001.
8. Yoon J., Jarrett D., Van der Schaar M. Time-series generative adversarial networks. Advances in Neural Information Processing Systems. 2019. V. 32. P. 5508–5518.
9. Eckerli F., Osterrieder J. Generative adversarial networks in finance. arXiv preprint. 2021. arXiv:2106.06364.
10. Eskandari N.M.R., Hamdi S.M., Filali B.S. SeriesGAN: Time series generation via adversarial and autoregressive learning. arXiv preprint. 2024. arXiv:2410.21203.
11. Seol S., Lee J., Yoon J., Kim B. Improving SOH estimation for lithium-ion batteries using TimeGAN. Machine Learning: Science and Technology. 2023. V. 4. № 4. P. 045007.
12. Roque L., Soares C., Cerqueira V., Torgo L. Latent generative transformer augmentation: Controlled synthetic generation of time series via variational multi-head attention. arXiv preprint. 2025. arXiv:2507.23615.
13. Zhang H., Wang L., Chen Y., Li J. Time series data augmentation for energy consumption data based on improved TimeGAN. Sensors. 2024. V. 25. № 2. P. 493.
14. Challu C., Olivares K.G., Oreshkin B.N. et al. N-HiTS: Neural hierarchical interpolation for time series forecasting. Proceedings of the AAAI Conference on Artificial Intelligence. 2023. V. 37. № 6. P. 6989–6997.
15. Talagala T.S., Hyndman R.J., Athanasopoulos G. Meta-learning how to forecast time series. Journal of Forecasting. 2023. V. 42. № 6. P. 1476–1501.
16. Norton D.A., Ott E., Pomerance A. et al. Tailored forecasting from short time series via meta-learning. arXiv preprint. 2025. arXiv: 2501.16325.
17. Wang X., Hyndman R.J., Li F., Kang Y. Forecast combinations: An over 50-year review. International Journal of Forecasting. 2023. V. 39. № 4. P. 1518–1547.
18. Goodfellow I., Pouget-Abadie J., Mirza M. et al. Generative adversarial nets. Advances in Neural Information Processing Systems. 2014. V. 27. P. 2672–2680.
19. Radford A., Metz L., Chintala S. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint. 2015. arXiv:1511.06434.
20. Salimans T., Goodfellow I., Zaremba W. et al. Improved techniques for training GANs. Advances in Neural Information Processing Systems. 2016. V. 29. P. 2234–2242.
21. Karras T., Aila T., Laine S., Lehtinen J. Progressive growing of GANs for improved quality, stability, and variation. arXiv preprint. 2017. arXiv:1710.10196.
22. Solycheva O.M. Lipshitsevy operatory superpozitsii na metricheskikh polugruppakh i abstraktnykh vypuklykh konusakh otobrazhenij konechnoj L-variatsii. Sibirskij matematicheskij zhurnal. 2006. T. 47. № 3. S. 650–665. (in Russian)
23. Monitoring kachestva priema v VUZy [Elektronnyj resurs]. URL: https://ege.hse.ru/ (data obrashcheniya: 12.10.2024). (in Russian)
24. Esteban C., Hyland S.L., Rätsch G. Real-valued (medical) time series generation with recurrent conditional GANs. arXiv preprint. 2017. arXiv:1706.02633.
25. Chen R.T., Li X., Grosse R., Duvenaud D. Isolating sources of disentanglement in variational autoencoders. Advances in Neural Information Processing Systems. 2018. V. 31. P. 2610–2620.
26. Yoon J., Jarrett D., van der Schaar M. Time-series generative adversarial networks. Advances in Neural Information Processing Systems. 2019. V. 32. P. 5508–5518.
27. Brophy E., Wang Z., She Q., Ward T. Generative adversarial networks in time series: A systematic literature review. ACM Computing Surveys. 2023. V. 55. № 10. P. 1–31.
28. Forestier G., Petitjean F., Dau H.A. et al. Generating synthetic time series to augment sparse datasets. IEEE International Conference on Data Mining (ICDM). 2017. P. 865–870.