Error-feedback stochastic modeling strategy for time series forecasting with convolutional neural networks

Zhang, Xinze; He, Kun; Bao, Yukun

doi:10.1016/j.neucom.2021.06.051

Cited by 12 publications

(3 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In particular, Zhang et al. ( 2021 ) use 3 time series in total, a simulated AR(1) process, a bitcoin price series and an influenza-like illness series, to evaluate their non-parametric neural network method. While the influenza-like illness series may be a good forecasting case study, basically the same considerations as for exchange rates and stock prices hold for bitcoin prices, though bitcoin was presumably a less efficient market, especially in its infancy.…”

Section: Motivation and Common Pitfallsmentioning

confidence: 99%

Forecast evaluation for data scientists: common pitfalls and best practices

Hewamalage

Ackermann

Bergmeir

2022

Data Min Knowl Disc

View full text Add to dashboard Cite

Recent trends in the Machine Learning (ML) and in particular Deep Learning (DL) domains have demonstrated that with the availability of massive amounts of time series, ML and DL techniques are competitive in time series forecasting. Nevertheless, the different forms of non-stationarities associated with time series challenge the capabilities of data-driven ML models. Furthermore, due to the domain of forecasting being fostered mainly by statisticians and econometricians over the years, the concepts related to forecast evaluation are not the mainstream knowledge among ML researchers. We demonstrate in our work that as a consequence, ML researchers oftentimes adopt flawed evaluation practices which results in spurious conclusions suggesting methods that are not competitive in reality to be seemingly competitive. Therefore, in this work we provide a tutorial-like compilation of the details associated with forecast evaluation. This way, we intend to impart the information associated with forecast evaluation to fit the context of ML, as means of bridging the knowledge gap between traditional methods of forecasting and adopting current state-of-the-art ML techniques.We elaborate the details of the different problematic characteristics of time series such as non-normality and non-stationarities and how they are associated with common pitfalls in forecast evaluation. Best practices in forecast evaluation are outlined with respect to the different steps such as data partitioning, error calculation, statistical testing, and others. Further guidelines are also provided along selecting valid and suitable error measures depending on the specific characteristics of the dataset at hand.

show abstract

Section: Motivation and Common Pitfallsmentioning

confidence: 99%

Forecast evaluation for data scientists: common pitfalls and best practices

Hewamalage

Ackermann

Bergmeir

2022

Data Min Knowl Disc

View full text Add to dashboard Cite

show abstract

“…These applications are nevertheless not exclusively focused on addressing the trade-off between locality and globality as we define it, but rather in intersection of RL and DL methods as a whole (e.g. employing RL mechanisms to tune the parameters of Generative Adversarial Networks [GAN] for replacing the sequential nature of RNN's in TS predictions - (Zhang et al, 2021)).…”

Section: Co-authorship Analysismentioning

confidence: 99%

“…In the work from (Zhang et al, 2021), we find a NN-based solution inspired by a policylike strategy (thus slightly hinting at the convergence of the research streams we proposed here) by adaptively building the network through iterative construction. The authors do this by proposing an error-feedback stochastic configuration strategy for designing a CNN (benefiting from the same performance on high-dimensionality shown by (Sen et al, 2019).…”

Section: Global-local Frameworkmentioning

confidence: 99%

Intersecting reinforcement learning and deep factor methods for optimizing locality and globality in forecasting: a review

Sousa

Henriques

2023

Preprint

View full text Add to dashboard Cite

In the current context of big data, the operational forecasting problems are more and more frequently involving the prediction of collections of multivariate, high-dimensional, related time series. The challenge of forecasting groups of time series can be tackled by fitting a single function to all series (global approach) or assuming each series to be a separate prediction problem and fitting one function to each problem (local approach). Although some global models show promising results against local benchmarks, there is still unclarity on how to best control generalization and structural assumptions by calibrating the exploitation of global patterns with local components. This is heavily reliant on the level of complexity from the dependency relationships exhibited between the series within a collection (e.g. indirect latent, local covariate relationships or noise effects) and other dimensions, such as the heterogeneity of the series and its length. For these reasons, there is an acknowledged need to invest further in developing scalable and data-driven hybrid models. The literature proposes Deep Learning (DL) based models built from different data generating processes (such as Auto-Regressive) which combine global models with classical local models (global-local frameworks). These methods frequently capture globality through factorization models or latent deep components and locality via classical local models or networks. These methods present non-linear modelling capabilities but are still either limited by solid theoretical assumptions (e.g. linearity of the underlying data) or predominantly one-dimensional, i.e., the future predictions for a single dimension mainly depend on past values from that same dimension. Simultaneously, dynamic Reinforcement Learning (RL) based models have been widely explored to optimize the balance of global and local signals in general regression problems, frequently through Q-learning algorithms. These derive from the equilibrium of cooperative strategies or the societal value in multi-agent learning systems. Similar models have been successfully applied to time series forecasting problems (such as stock market predictions), addressing the risk of method generalization under varying conditions. This paper conducted a concise literature review focused on these two research streams (global-local frameworks and RL based models) to optimize the balance between globality and locality in forecasting collections of time series. It focuses on their evolution across time and hints at opportunities to close some of the research gaps by intersecting both propositions. We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and achieved a selection of 118 publications since 2000. The main findings revealed that global models have achieved strong expressiveness in capturing the most complex structural patterns while still enabling probabilistic outcomes to be delivered through uncertainty estimates. On the other hand, RL based methods depict great benefits in mitigating the risks of generalization by imprinting contextual diversity when predicting each step ahead for each series. Within those, the adoption of other computational learning or evolutionary-based methods (e.g. Genetic Programming) to improve the parametrization of the learning policies is also highlighted as an area of future work yet to be uncovered.

show abstract