Skip to main content

Advertisement

NQF-RNN: probabilistic forecasting via neural quantile function-based recurrent neural networks

  • Published:
Applied Intelligence Aims and scope Submit manuscript

Abstract

Probabilistic forecasting offers insights beyond point estimates, supporting more informed decision-making. This paper introduces the Neural Quantile Function with Recurrent Neural Networks (NQF-RNN), a model for multistep-ahead probabilistic time series forecasting. NQF-RNN combines neural quantile functions with recurrent neural networks, enabling applicability across diverse time series datasets. The model uses a monotonically increasing neural quantile function and is trained with a continuous ranked probability score (CRPS)-based loss function. NQF-RNN’s performance is evaluated on synthetic datasets generated from multiple distributions and six real-world time series datasets with both periodicity and irregularities. NQF-RNN demonstrates competitive performance on synthetic data and outperforms benchmarks on real-world data, achieving lower average forecast errors across most metrics. Notably, NQF-RNN surpasses benchmarks in CRPS, a key probabilistic metric, and tail-weighted CRPS, which assesses tail event forecasting with a narrow prediction interval. The model outperforms other deep learning models by 5% to 41% in CRPS, with improvements of 5% to 53% in left tail-weighted CRPS and 6% to 34% in right tail-weighted CRPS. Against its baseline model, DeepAR, NQF-RNN achieves a 41% improvement in CRPS, indicating its effectiveness in generating reliable prediction intervals. These results highlight NQF-RNN’s robustness in managing complex and irregular patterns in real-world forecasting scenarios.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save

Springer+ Basic
$34.99 /Month
  • Get 10 units per month
  • Download Article/Chapter or eBook
  • 1 Unit = 1 Article or 1 Chapter
  • Cancel anytime
Subscribe now

Buy Now

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Algorithm 1
Fig. 3
Fig. 4

Similar content being viewed by others

Explore related subjects

Discover the latest articles, news and stories from top researchers in related subjects.

Data Availability

All datasets can be downloaded in the following URLs: − synthetic data

(https://doi.org/10.7910/DVN/W04WWC) − electricity [1, 44]

(https://archive.ics.uci.edu/dataset/321/electricityloaddiagrams20112014) − traffic [1, 44]

(https://archive.ics.uci.edu/dataset/204/pems+sf) − solar [45]

(https://www.nrel.gov/grid/solar-power-data.html) − M4-hourly [46]

(https://github.com/Mcompetitions/M4-methods/tree/master) − tourism-monthly, tourism-quarterly [47]

(https://robjhyndman.com/publications/the-tourism-forecasting-competition)

References

  1. Salinas D, Flunkert V, Gasthaus J, Januschowski T (2020) Deepar: Probabilistic forecasting with autoregressive recurrent networks. Int J Forecast 36(3):1181–1191

    Article  Google Scholar 

  2. Rasul K, Sheikh A-S, Schuster I, Bergmann UM, Vollgraf R (2021) Multivariate probabilistic time series forecasting via conditioned normalizing flows. In: International conference on learning representations. https://openreview.net/forum?id=WiGQBFuVRv

  3. Yang L, Zhang Z, Song Y, Hong S, Xu R, Zhao Y, Zhang W, Cui B, Yang M-H (2023) Diffusion models: A comprehensive survey of methods and applications. ACM Comput Surv 56(4):1–39

    Article  MATH  Google Scholar 

  4. Cannon AJ (2011) Quantile regression neural networks: Implementation in r and application to precipitation downscaling. Comput Geosci 37(9):1277–1284

    Article  MATH  Google Scholar 

  5. Gasthaus J, Benidis K, Wang Y, Rangapuram SS, Salinas D, Flunkert V, Januschowski T (2019) Probabilistic forecasting with spline quantile function rnns. In: The 22nd international conference on artificial intelligence and statistics, pp 1901–1910. PMLR

  6. Chilinski P, Silva R (2020) Neural likelihoods via cumulative distribution functions. In: Conference on uncertainty in artificial intelligence, pp 420–429. PMLR

  7. Alcántara A, Galván IM, Aler R (2022) Direct estimation of prediction intervals for solar and wind regional energy forecasting with deep neural networks. Eng Appl Artif Intell 114:105128

    Article  MATH  Google Scholar 

  8. Sun R, Li C-L, Arik SÖ, Dusenberry MW, Lee C-Y, Pfister T (2023) Neural spline search for quantile probabilistic modeling. In: Proceedings of the AAAI conference on artificial intelligence, vol 37, pp 9927–9934

  9. Wen R, Torkkola K, Narayanaswamy B, Madeka D (2017) A multi-horizon quantile recurrent forecaster. In: NIPS 2017 time series workshop

  10. Zhang X-Y, Watkins C, Kuenzel S (2022) Multi-quantile recurrent neural network for feeder-level probabilistic energy disaggregation considering roof-top solar energy. Eng Appl Artif Intell 110:104707

    Article  MATH  Google Scholar 

  11. Gouttes A, Rasul K, Koren M, Stephan J, Naghibi T (2021) Probabilistic time series forecasting with implicit quantile networks. In: ICML 2021 Time Series Workshop

  12. Park Y, Maddix D, Aubet F-X, Kan K, Gasthaus J, Wang Y (2022) Learning quantile functions without quantile crossing for distribution-free time series forecasting. In: International conference on artificial intelligence and statistics, pp 8127–8150 . PMLR

  13. Hu J, Tang J, Liu Z (2024) A novel time series probabilistic prediction approach based on the monotone quantile regression neural network. Inf Sci 654:119844

    Article  MATH  Google Scholar 

  14. Cannon AJ (2018) Non-crossing nonlinear regression quantiles by monotone composite quantile regression neural network, with application to rainfall extremes. Stoch Env Res Risk A 32:3207–3225

    Article  MATH  Google Scholar 

  15. Koenker R, Hallock KF (2001) Quantile regression. J Econ Perspect 15(4):143–156

    Article  MATH  Google Scholar 

  16. Wang Q, Ma Y, Zhao K, Tian Y (2020) A comprehensive survey of loss functions in machine learning. Annals of Data Science, pp 1–26

  17. Zou H, Yuan M (2008) Composite quantile regression and the oracle model selection theory. Ann Stat 36(3):1108–1126. Accessed 2024-03-16

  18. Moon SJ, Jeon J-J, Lee JSH, Kim Y (2021) Learning multiple quantiles with neural networks. J Comput Graph Stat 30(4):1238–1248

    Article  MathSciNet  MATH  Google Scholar 

  19. Gneiting T, Raftery AE (2007) Strictly proper scoring rules, prediction, and estimation. J Am Stat Assoc 102(477):359–378

    Article  MathSciNet  MATH  Google Scholar 

  20. Rasul K, Seward C, Schuster I, Vollgraf R: Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In: International conference on machine learning, pp 8857–8868 (2021). PMLR

  21. Zhang H, Zhang Z (1999) Feedforward networks with monotone constraints. In: IJCNN’99. international joint conference on neural networks. proceedings (Cat. No. 99CH36339), vol 3, pp 1820–1823 . IEEE

  22. Friederichs P, Thorarinsdottir TL (2012) Forecast verification for extreme value distributions with an application to probabilistic peak wind prediction. Environmetrics 23(7):579–594

    Article  MathSciNet  MATH  Google Scholar 

  23. Lin F, Zhang Y, Wang K, Wang J, Zhu M (2022) Parametric probabilistic forecasting of solar power with fat-tailed distributions and deep neural networks. IEEE Trans Sustain Energy 13(4):2133–2147

    Article  MATH  Google Scholar 

  24. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780

    Article  MATH  Google Scholar 

  25. Jozefowicz R, Zaremba W, Sutskever I (2015) An empirical exploration of recurrent network architectures. In: International conference on machine learning, pp 2342–2350. PMLR

  26. Sutskever I, Vinyals O, Le QV (2014) Sequence to sequence learning with neural networks. Advances in neural information processing systems 27

  27. Goodfellow I, Bengio Y, Courville A (2016) Deep Learning. MIT Press. http://www.deeplearningbook.org

  28. Williams RJ, Zipser D (1989) A learning algorithm for continually running fully recurrent neural networks. Neural Comput 1(2):270–280

    Article  MATH  Google Scholar 

  29. Rudin W, et al (1976) Principles of Mathematical Analysis vol 3. McGraw-hill

  30. Koenker R, Machado JA (1999) Goodness of fit and related inference processes for quantile regression. J Am Stat Assoc 94(448):1296–1310

    Article  MathSciNet  MATH  Google Scholar 

  31. Taylor JW (1999) Evaluating volatility and interval forecasts. J Forecast 18(2):111–128

    Article  MATH  Google Scholar 

  32. Davis PJ, Rabinowitz P (2007) Methods of numerical integration. Courier Corporation

  33. Van Der Walt S, Colbert SC, Varoquaux G (2011) The numpy array: a structure for efficient numerical computation. Comput Sci Eng 13(2):22–30

    Article  MATH  Google Scholar 

  34. Gneiting T, Ranjan R (2011) Comparing density forecasts using threshold-and quantile-weighted scoring rules. J Bus Econ Stat 29(3):411–422

    Article  MathSciNet  MATH  Google Scholar 

  35. Zamo M, Naveau P (2018) Estimation of the continuous ranked probability score with limited information and applications to ensemble weather forecasts. Math Geosci 50(2):209–234

  36. Bollerslev T (1986) Generalized autoregressive conditional heteroskedasticity. J Econ 31(3):307–327

    Google Scholar 

  37. Meinshausen N, Ridgeway G (2006) Quantile regression forests. J Mach Learn Res 7(6)

  38. Petropoulos F, Apiletti D, Assimakopoulos V, Babai MZ, Barrow DK, Taieb SB, Bergmeir C, Bessa RJ, Bijak J, Boylan JE et al (2022) Forecasting: theory and practice. Int J Forecast 38(3):705–871

    Article  Google Scholar 

  39. Hewamalage H, Bergmeir C, Bandara K (2021) Recurrent neural networks for time series forecasting: Current status and future directions. Int J Forecast 37(1):388–427

    Article  MATH  Google Scholar 

  40. Kunz M, Birr S, Raslan M, Ma L, Januschowski T (2023) Deep learning based forecasting: a case study from the online fashion industry. In: Forecasting with artificial intelligence: theory and applications, pp 279–311. Springer

  41. Chung J, Gulcehre C, Cho K, Bengio Y (2014) Empirical evaluation of gated recurrent neural networks on sequence modeling. In: NIPS 2014 workshop on deep learning, December 2014

  42. Challu C, Olivares KG, Oreshkin BN, Ramirez FG, Canseco MM, Dubrawski A (2023) Nhits: Neural hierarchical interpolation for time series forecasting. Proceedings of the AAAI conference on artificial intelligence 37:6989–6997

    Article  Google Scholar 

  43. Chen Z, Ma M, Li T, Wang H, Li C (2023) Long sequence time-series forecasting with deep learning: A survey. Inf Fusion 97:101819

    Article  MATH  Google Scholar 

  44. Yu H-F, Rao N, Dhillon IS (2016) Temporal regularized matrix factorization for high-dimensional time series prediction. Advances in neural information processing systems 29

  45. Lai G, Chang W-C, Yang Y, Liu H (2018) Modeling long-and short-term temporal patterns with deep neural networks. In: The 41st international ACM SIGIR conference on research & development in information retrieval, pp 95–104

  46. Makridakis S, Spiliotis E, Assimakopoulos V (2018) The m4 competition: Results, findings, conclusion and way forward. Int J Forecast 34(4):802–808

    Article  MATH  Google Scholar 

  47. Athanasopoulos G, Hyndman RJ, Song H, Wu DC (2011) The tourism forecasting competition. Int J Forecast 27(3):822–844

    Article  MATH  Google Scholar 

Download references

Funding

This work was supported by the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF-2023S1A5A2A03086550).

Author information

Authors and Affiliations

Authors

Contributions

Jungyoon Song: Conceptualization, Methodology, Visualization, Writing − original draft, Woojin Chang: Writing − review & editing, Supervision, Jae Wook Song: Conceptualization, Writing − review & editing, Validation, Supervision.

Corresponding author

Correspondence to Jae Wook Song.

Ethics declarations

Ethical Approval

All datasets used in this article are publicly available, and no consent was required for their use.

Conflicts of Interest

The authors have no competing interests to declare that are relevant to the content of this article.

Competing interests

The authors have no financial interests that could have appeared to influence the work reported in this article.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendices

Appendix A Abbreviations & Symbols

The abbreviations and symbols used in this paper are summarized in Table 9.

Table 9 List of abbreviations & symbols

Appendix B Proper scoring rules

Based on the work of [19],

$$\begin{aligned} S(P,Q) = \int S(P,\omega )\, dQ(\omega ) \end{aligned}$$
(B1)

for the expected score under Q when the probabilistic forecast is P. The scoring rule S is proper relative to \(\mathcal {P}\) if

$$\begin{aligned} S(Q, Q) \ge S(P,Q) \text { for all } P,Q \in \mathcal {P} \end{aligned}$$
(B2)

It is strictly proper relative to \(\mathcal {P}\) if (B2) holds with equality if and only if \(P = Q\).

Appendix C Detailed preprocessing of datasets

1.1 C.1 Synthetic datasets

The synthetic dataset used in this study comprises four distinct time series input variables. The first input variable consists of scaled observations from a one-lagged time series, denoted by \(z_{i,t-1}\). The second input variable serves as a time index, represented by \(t-1\) within the range \([0, \dots , 41]\). The third input variable represents weekly seasonality, calculated as \(t - 1 \;\text {mod}\; 7\), within the range \([0, \dots , 6]\). The fourth input variable is an item-related covariate, represented by an integer identifier that specifies the group to which a particular observation belongs. To standardize these input features, both the second and third variables are normalized using z-score normalization.

1.2 C.2 Real-world datasets

For the real-world datasets, each time series input is denoted as \(x_{i,t} \in \mathbb {R}^{T \times D}\), where supplementary input variables are associated with time-dependent covariates. The first input variable is the scale-adjusted observation of a one-lagged time series, \(z_{i,t-1}\). Additional input variables include time-related covariates derived from the date information in each dataset, such as hours (ranging from 0 to 23), weekdays (ranging from 0 to 6), months (ranging from 1 to 12), and quarters (ranging from 1 to 4). These covariates are represented by integer values and used as input features.

Another time-related covariate, age, represents the temporal distance from the initial observation within each time series. This variable is adjusted according to the time granularity specific to each dataset.

Additionally, an item-related covariate is included to identify specific series within each dataset. This covariate is also represented by an integer value; for example, in the electricity dataset, it corresponds to the customer index (ranging from 0 to 370) and is linked to the input embedding dimension. The time-related covariates undergo z-score normalization to ensure consistency across varied scales and to enhance model performance. The dataset is generated using a stride equal to the decoder length to avoid overlapping instances in model training, validation, and testing during forecasting.

Appendix D Hyperparameter sets for models

1.1 D.1 Synthetic datasets

For the synthetic datasets, the hyperparameters for each model are defined as follows. For the QRF model, the hyperparameters include the number of estimators, set to either 50 or 100, and the maximum depth, set within the range of 5, 10, or 15. Deep learning-based models use a standardized set of hyperparameters, including a learning rate of either 0.005 or 0.01 and item-related embedding dimensions of 1 or 5. The QRNN, SQF-RNN, IQN-RNN, and NQF-RNN models are configured with RNN hidden dimensions set to 20 or 60, with three hidden layers. The SQF-RNN model specifies the number of knot positions L as 5 or 10. For the NQF-RNN model, the quantile function comprises DNN layers configured as [32, 16] or [8, 4]. The NHITS model incorporates hyperparameters, including an aggregation kernel size defined as [2, 2, 1], [2, 2, 2], or [4, 2, 1], and frequency downsampling set to [4, 2, 1], [14, 7, 1], or [28, 14, 1]. This model also utilizes three blocks, with linear interpolation mode and Maxpool1d as the pooling mode.

1.2 D.2 Real-world datasets

For the real-world datasets, the hyperparameters for each model are as follows. The QRF model includes the number of estimators, set to either 50 or 100, and the maximum depth, set within the range of 5, 10, or 15. Deep learning-based models employ a standard set of hyperparameters, including a learning rate of 0.001 or 0.005 and item-related embedding dimensions of 5 or 20. The QRNN, SQF-RNN, IQN-RNN, and NQF-RNN models have RNN hidden dimensions set to 20, 40, or 60, across three hidden layers. The SQF-RNN model specifies a number of knot positions L as 5 or 10. For the NQF-RNN model, the quantile function incorporates DNN layers configured as [64, 32, 16, 8], [128, 64], [64, 32], or [16, 8]. The NHITS model includes hyperparameters such as an aggregation kernel size set to [2, 2, 1], [2, 2, 2], or [4, 2, 1], and frequency downsampling set to [4, 2, 1], [24, 12, 1], or [168, 24, 1]. This model also employs three blocks, using linear interpolation mode and Maxpool1d as the pooling mode. For each hyperparameter set, the model achieving optimal validation performance is selected for the final evaluation.

Appendix E Visualization of probabilistic forecasting in four real-world datasets

Examples of results for the traffic, M4-hourly, tourism-monthly, and tourism-quarterly datasets are presented in Figs. E1, E2, E3, and E4, respectively.

Fig. 5
figure 5

Visualization of probabilistic forecasts for traffic

Fig. 6
figure 6

Visualization of probabilistic forecasts for M4-hourly

Fig. 7
figure 7

Visualization of probabilistic forecasts for tourism-monthly

Fig. 8
figure 8

Visualization of probabilistic forecasts for tourism-quarterly

In panel (a), the solid gray line positioned outside the yellow box represents the conditioning range, which serves as input to the models along with various covariates. Using this input, the models perform multistep-ahead forecasting for the prediction range, depicted by a solid black line. Panels (b) through (j) display the probabilistic forecasting results within the prediction range for different models. The solid red line denotes the point forecasts, while the dotted red, orange, and yellow lines correspond to the quantiles of 0.1/0.9, 0.05/0.95, and 0.01/0.99 of the prediction intervals, respectively.

Figure E1 suggests that the time series across models exhibit comparable patterns but differ in scale of \(y-axis\). Figure E2 shows that, while the time series generally maintain consistent patterns, notable scale differences are also present. Figure E3 indicates that the time series share similar scales but show irregularities. In Fig. E4, the time series predominantly adhere to consistent historical patterns.

In terms of performance, the GARCH and QRF models demonstrate inadequate predictive accuracy. QRNN encounters challenges in forecasting beyond the median range, while DeepAR tends to produce relatively broader prediction intervals compared to other probabilistic time series models. In contrast, other deep learning-based probabilistic time series models exhibit more effective pattern learning and generate reasonable prediction intervals. Notably, NQF-RNN achieves the narrowest prediction intervals across all datasets, effectively capturing the evolution of the original time series.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Song, J., Chang, W. & Song, J.W. NQF-RNN: probabilistic forecasting via neural quantile function-based recurrent neural networks. Appl Intell 55, 183 (2025). https://doi.org/10.1007/s10489-024-06077-7

Download citation

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1007/s10489-024-06077-7

Keywords