A new automatic forecasting method based on explainable deep dendritic artificial neural network Authors: Eren Bas, Erol Egrioglu Published: 2025 ABSTRACT Forecasting practitioners need automated forecasting methods that do not involve subjective decisions. For this reason, in recent years, the literature has been trying to automate forecasting methods. Automated forecasting methods are usually based on various models or variable selection strategies and various hypothesis tests for data preprocessing. This study proposes input significance tests for model inputs of deep dendritic recurrent neural networks and model validity tests, in the context of explainability studies. The performance of the proposed tests is investigated with a simulation study. Secondly, based on the developed tests and some statistical tools, a new automatic forecasting method for deep dendritic recurrent neural networks is proposed. The forecasting ability of the proposed method is compared with other successful forecasting methods in the literature using M3 and M4 competition time series. Introduction Knowing the future values of many time series encountered in real life plays an important role for decision makers in terms of future planning. Forecasting methods are frequently needed in the fields of economics, finance, health, and engineering. Statistical forecasting methods focus on identifying the components of the time series and analysing them with statistical methods. In solving the forecasting problem, it is common to violate assumptions such as nonlinearity, normal distribution, and constant variance. Contemporary machine learning methods, such as artificial neural networks, which theoretically do not require assumptions such as normality, linearity, and constant variance, have become popular for the problem of forecasting. Artificial neural networks can easily adapt to non-linear relationships between variables due to their flexible model structure. Deep neural networks can be divided into recurrent and feed-forward networks. Since recurrent neural networks include time steps in their architecture and formulas, they can provide a model suitable for the forecasting of time series. For this reason, they are frequently used in the literature to obtain future forecasts. In the literature, deep learning methods have found a place in many forecasting studies1. conducted a literature study on time series forecasting with deep learning. This is a good compilation of the work carried out up to 20222. focused on deep learning methods for temperature forecasting, and Stochastic Adversarial Video Prediction (SAVP) and a convolutional neural network (CNN) are the deep learning models considered in the research3. presented a deep learning method for power load forecasting4. present an integrated method for volatility forecasting using long short-term memory (LSTM), gated recurrent unit (GRU) deep feedforward neural networks, and reinforcement learning Q-learning algorithm5. proposed a hybrid deep neural network architecture for sales forecasting by connecting LSTM and CNN deep neural networks in parallel and series6. presented an algorithm based on leveraging quantum-inspired chimp optimization and deep neural networks for profit forecasting7. proposed a new hybrid deep artificial neural network for the forecasting of air pollution data8. proposed a deep feedback artificial neural network model based on the Extreme Gradient Boosting variable selection strategy for electricity spot-price forecasting9. introduced a hybrid approach of CNN and LSTM networks to predict the air quality index10. proposed a deep neural network with an attention mechanism for wind power forecasting11. proposed a new deep neural network: “DeepDenT”12. presented a hybrid artificial neural network method with deep feedback. Despite the popularity of deep neural networks, there is a problem that these networks contain too many hyperparameters, and the automatic determination of these hyperparameters. In the literature, there are automatic forecasting methods designed for some types of neural networks and classical forecasting methods, but not for deep neural networks13. introduced a new automatic forecasting model based on the ARIMA model. Different approaches are used to identify and eliminate non-stationarity in the method14. presented an automatic forecasting algorithm for general regression neural network models15. introduced an automated ensemble approach to classical forecasting methods16. presented an ensemble approach of many automatic forecasting methods. In this method, the use of both machine learning and classical methods brings a significant advantage17. introduced a generalized version of Theta, the automatic forecasting method proposed by18, and tested it on M4 data19. compared the performance of automatic forecasting methods using the results of the 2018 Swedish general election data. It is determined that the ETS method is the most successful method for this data set20. presented an automatic forecasting for electricity demand based on random forest and some other machine learning methods. The method works based on data mining algorithms21. presented an automatic forecasting method based on a multilayer perceptron that automatically identifies the components in the time series and works in NARX logic21. presented an automatic forecasting method based on a multilayer perceptron that automatically identifies the components in the time series and works in NARX logic. Although statistical tests are used to determine the time series components in the method, any testing of model inputs and explainable artificial intelligence is excluded11. presented an automatic forecasting method based on a shallow artificial neural network. In11, model validity and input significance tests are the most critical points of the algorithm22. proposed automatic forecasting based on multiple models for forecasting Chickenpox in Hungary23. used various machine learning methods to automatically predict the eruption of a volcano in advance. However, this method is not exactly an automatic forecasting method, but rather a problem-specific determination of machine learning parameters24. presented an automated forecasting method for time series with long-run memory in univariate time series analysis. The algorithm works based on determining the most appropriate ARFIMA model25. proposed the “Timemixer++” method as an automatic forecasting method. Although there are many automated forecasting approaches in the literature, there is no work on automating deep neural networks. In this study, a new automatic forecasting method is presented by proposing an explainable version of the “DeepDenT” artificial neural network proposed by26. The first contribution of the study is the development of input significance tests and model validity tests for the DeepDenT neural network. In this context, in addition to testing each input in the explainable DeepDenT neural network, the importance factors of the inputs can also be calculated. The second contribution of the study is the introduction of an automatic forecasting method based on the explainable DeepDenT neural network. In the second section, information about the “DeepDenT” artificial neural network is given. Algorithms for model validity and input significance tests within the scope of explainable DeepDenT artificial neural network studies are presented, and the performance of these algorithms is demonstrated with a simulation study in the third section. In the fourth section, the proposed automatic forecasting algorithm is presented. The fifth section presents simulation study results. The sixth section presents the results obtained from the application of the proposed automatic forecasting algorithm on M3 and M4 competition data. In the last section, the findings are discussed, and information about future work is given. Related works This section provides a summary of technical information on some shallow and deep artificial neural networks in the literature that support the method proposed in this study. A shallow artificial neural network with a dendritic structure was proposed for the first time in the study by27, which was developed by modifying the mathematical model of neurons in artificial neural networks. In a dendritic artificial neural network, outputs are generated through nonlinear transformations of inputs and additive and multiplicative combination operations. Equations (1–4) are used to calculate outputs in a dendritic artificial neural network.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{ij}=\frac{1}{1+\mathrm{e}\mathrm{x}\mathrm{p}(-k\left({w}_{ij}{Girdi}_{i}+{\theta\:}_{ij}\right))}\:\:i=\mathrm{1,2},\dots\:,p;j=\mathrm{1,2},\dots\:,q$$\end{document} The multiplication of transformed inputs is made for each dendrite by using the dendrite function, which is given below.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{j}=\left(\prod\:_{i=1}^{p}{Y}_{ij}\right);j=\mathrm{1,2},\dots\:,q$$\end{document} All outputs of the dendrites are combined by using the following membrane function3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:V=\sum\:_{j=1}^{q}{Z}_{j}$$\end{document} Finally, outputs of the DNM-ANN are calculated by transforming of membrane outputs in Eq. (4).4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:output=\frac{1}{1+exp\left(-{k}_{soma}(V-\:{\theta\:}_{soma})\right)}$$\end{document} The success of the dendritic artificial neural network in solving forecasting problems suggests the creation of a deep artificial neural network inspired by this cell structure. In addition to the success of deep recurrent neural networks in solving the forecasting problem in the literature, deep neural networks with cell structures such as LSTM and GRU have also been found to be preferable for the forecasting problem in recent years. A simple deep feedforward neural network is a deep artificial neural network that operates with a cellular structure and has a hierarchical structure that provides a feedback mechanism, while possessing mathematical operations similar to those of a multi-layer perceptron. In this deep artificial neural network, the outputs are obtained by sequentially applying the (5), (6), and (7) formulas in the output calculation.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-k}^{1}=\sigma\:\left({X}_{t-k}{W}^{1}+{h}_{t-k-1}{R}^{1}+{b}^{1}\right)=f\left({h}_{t-k-1}^{1},{X}_{t-k},{{\uptheta\:}}^{1}\right),\:k=\mathrm{1,2},\dots\:,h$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-k}^{j}=\sigma\:\left({h}_{t-k}^{j-1}{W}^{j}+{h}_{t-k-1}^{j}{R}^{j}+{b}^{j}\right)=f\left({h}_{t-k-1}^{j},{h}_{t-k}^{j-1},{{\uptheta\:}}^{j}\right),\:k=\mathrm{1,2},\dots\:,h,\:j=2,\dots\:,m$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{y}_{t}}=\sigma\:\left({V}_{FC}{h}_{t-1}^{m}+{b}_{FC}\right)$$\end{document} By converting the dendritic neuron structure into a cellular structure, it is possible to create a deep artificial neural network. Considering the methods that inspired this deep artificial neural network, it can be expected to be successful in the forecasting problem26. introduced a new deep feedback artificial neural network called “DeepDenT”. This neural network is formed by combining dendritic cells (DnC) in a certain hierarchical structure. The DnC cell has an architecture that works with multiple inputs and produces multiple outputs. The calculation details and calculation formulas of DnC are given in26. The DeepDenT neural network combines DnCs. DeepDenT is a deep neural network created by connecting dendritic cells sequentially according to time and contains a fully connected structure in the output layer. The inputs of DeepDent are the lagged variables of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t}=({y}_{t},{y}_{t-1,\dots\:,}{y}_{t-p+1})$$\end{document} according to the number of hidden layer nodes. The output of the DeepDent deep recurrent neural network is a one-step-ahead forecast of the time series. The architecture has h time steps, q hidden layers, m dendrites, and p inputs. DnCs in the same hidden layer have equal weights and side values. Many parameters are used in the calculations, and in order to overcome this notational complexity, the common parameters of each hidden layer of the network are combined as given in (8). In DeepDenT, the parameters vary from hidden layer to hidden layer, but they stay the same for the same hidden layer units, i.e., time steps, and they’re shared.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Theta\:}}^{j}=\left[{{W}_{x}}^{j},{\theta\:}_{x}^{j},{{W}_{h}}^{j},{\theta\:}_{h}^{j},{k}_{soma}^{j},{\theta\:}_{soma}^{j},{k}^{j}\right),\:j=\mathrm{1,2},\dots\:,q$$\end{document} Equation (9) is used to calculate the outputs from the first hidden layer. The function f is given here as a general notation.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-k}^{1}=f\left({h}_{t-k-1}^{1},{x}_{t-k},{{\Theta\:}}^{1}\right),\:k=\mathrm{1,2},\dots\:,h$$\end{document} In Eq. (9), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-k}^{1}$$\end{document} denotes the output of the first hidden layer unit in the first hidden layer. For the other hidden layer units, Eq. (10) is used. As seen in Eq. (10), the inputs of the cells are no longer lagged variables, but signals from the previous hidden layer.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-k}^{j}=f\left({h}_{t-k-1}^{j},{h}_{t-k}^{j-1},{{\Theta\:}}^{j}\right),\:k=\mathrm{1,2},\dots\:,h,\:j=2,\dots\:,q$$\end{document} When calculating the outputs of DnCs, the order of calculation is important and always starts from the left-hand side. The output of DeepDenT is calculated using formula (11) based on the output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-1}^{q}$$\end{document} of the cell shown in bold.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}=\sigma\:\left({W}_{FC}{h}_{t-1}^{q}+{b}_{FC}\right)$$\end{document} 26 present a successful algorithm based on the differential evolution algorithm for training DeepDenT. Explainable deep dendritic artificial neural network Explainability studies in artificial neural networks have gained importance in recent years. There are many different explainability studies in the literature. In general, explainability is to tell the story of how the output of the model is generated, given an input. In classical forecasting models, since the models are linear, explainability can be realized directly by interpreting the model coefficients. In the classical forecasting model, the significance of each variable and model validity are realised through hypothesis tests. In this study, model validity and input significance tests for the “DeepDenT” deep feedback neural network for explainability are presented for the first time, inspired by the method developed by28 for a multilayer perceptron. A multiple regression method for the outputs of the trained neural network is used to apply the tests and determine the importance scores of the inputs. For explainable DeepDenT, model validity, input significance, and importance scores are obtained in the following algorithm. Algorithm 1: Tests for Explainable DeepDenT. Step 1. The DeepDenT neural network is trained. Step 2. The outputs of the trained DeepDent are calculated for each input while the averages of the other inputs are fixed. From the trained DeepDent, the prediction function is obtained by Eq. (12).12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{x}}_{t}^{}=g\left({x}_{t-1},\dots\:,{x}_{t-p}\:\:;\:\widehat{W}\:,\:\:\widehat{B}\right)$$\end{document} Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:nts$$\end{document} be the number of learning samples. The outputs of inputs are denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Y}}_{j}^{\left(k\right)}$$\end{document} ,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j=\mathrm{1,2},...,nts,k=\mathrm{1,2},...,p$$\end{document} and performed Eq. (13). When calculating outputs for any input, other inputs are fixed at their arithmetic mean.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Y}}_{j}^{\left(k\right)}=g\left({x}_{t-1}=mean\left({x}_{t-1}\right),\dots\:,{x}_{t-k-1}=mean\left({x}_{t-k-1}\right),{x}_{t-k},{x}_{t-k+1}=mean\left({x}_{t-k+1}\right),\dots\:,{x}_{t-p}=mean\left({x}_{t-p}\right)\right)\:)$$\end{document} Step 3. A linear model is created and predict between the outputs obtained and the target values. In the model, the targets are employed as the response variable and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Y}}_{j}^{\left(k\right)}$$\end{document} are used as covariates. The linear model is given by Eq. (14), and it is the full rank linear model.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t}={\beta\:}_{0}+{\sum\:}_{k=1}^{p}{\beta\:}_{k}{\widehat{Y}}^{\left(k\right)}+{\epsilon\:}_{t}$$\end{document} The absolute values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\beta\:}_{k}}\:$$\end{document}estimated from this model are interpreted as the importance score of each input. Step 4. Perform partial F-tests to test. The hypothesis for the test is given below.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{0}:{\beta\:}_{1}={\beta\:}_{2}=\dots\:={\beta\:}_{k}=0\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{1}:\exists\:{\beta\:}_{k}\ne\:0,\:\:k=\mathrm{1,2},\dots\:,p$$\end{document}. Test statistics are calculated with Eq. (16).16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F=\frac{{\left(K\beta\:\right)}^{{\prime\:}}{\left[K{\left({X}^{{\prime\:}}X\right)}^{-1}K{\prime\:}\right]}^{-1}\left(K\beta\:\right)}{p{\widehat{\sigma\:}}^{2}}$$\end{document} In (16) K is the matrix of coefficients used to formulate hypotheses. The matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X$$\end{document} is the matrix of independent variables, and its ith row is given as follows:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{i}=\left[{1\:\widehat{Y}}^{\left(1\right)}\:{\widehat{Y}}^{\left(2\right)}\cdots\:{\widehat{Y}}^{\left(p\right)}\right];\:i=\mathrm{1,2},\dots\:,nts$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} represents the standard deviation estimate of the linear model.18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{\sigma\:}}^{2}=\frac{{Y}^{{\prime\:}}{\left[I-X{\left({X}^{{\prime\:}}X\right)}^{-1}X{\prime\:}\right]}^{-1}Y}{nts-p-1}$$\end{document} The distribution of the test statistic is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F(p,nts-p-1)$$\end{document}. The significance value is the probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P$$\end{document}(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F(p,nts-p-1)$$\end{document}>\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{F}_{value})$$\end{document}. Step 5. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\left(F\right(p,nts-p-1)>\:F\_value)<\alpha\:$$\end{document}, the DeepDent model is sufficient and proceeds to Step 6 for input significance tests. Otherwise, the model is inadequate, and the algorithm is stopped. Step 6. All inputs are tested with the following hypotheses.19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{0}:{\beta\:}_{k}=0\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{1}:{\beta\:}_{k}\ne\:0$$\end{document}. The test statistic is calculated with the following formula. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:K$$\end{document} is a matrix calculated according to the hypothesis. For example, if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}$$\end{document} is to be tested, the matrix is given by Eq. (20).20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:K=\left[\begin{array}{cccccc}0&\:1&\:0&\:0&\:0&\:0\end{array}\right]$$\end{document} The test statistic is calculated by using Eq. (16). The test statistic has \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F(1,nts-p-1)$$\end{document} distribution. The significance value is calculated with probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\left(F\right(1,nts-p-1)>\:{F}_{value})$$\end{document}. Step 7. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\left(F\right(1,nts-p-1)>\:{F}_{value})<\alpha\:$$\end{document}, the input is significant; otherwise, it is not significant. Step 8. Steps 6–7 are repeated for all inputs. In this algorithm, the critical values for the hypothesis tests applied in Step 4 and Step 6 are obtained from the F distribution. However, this distribution is usable under the assumption that the distribution of errors is normal. Therefore, this step has led to the use of the Bootstrap hypothesis test, which does not require any assumption control. Although the algorithm is the same, the algorithm for obtaining the critical values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F(p,nts-p-1)$$\end{document} in Step 4 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F(1,nts-p-1)$$\end{document} in Step 7 using the bootstrap method is given below. Algorithm 2: Model adequacy, input significance, bootstrap hypothesis tests, and significance scores for explainable DeepDent. Steps 1,2, and 3. It is applied in the same way as the steps in Algorithm 1. Step 4. Perform partial bootstrap F-tests to test model adequacy. The hypothesis for the test is given in (15). Test statistics are calculated with the (16) Equation. The critical value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Fbootstrap$$\end{document} for the decision stage is obtained with the bootstrap approach. For this purpose, a bootstrap sample is drawn from the learning samples created for the model given in (13) with the substitution condition, and the F value given in Eq. (16) is calculated for each bootstrap sample. For example, a bootstrap sample can be selected for each predictor in the model as shown below. 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {\hat{Y}_{{10}}^{{\left( 1 \right)}} ,\hat{Y}_{2}^{{\left( 1 \right)}} ,\hat{Y}_{3}^{{\left( 1 \right)}} , \ldots ,\hat{Y}_{{10}}^{{\left( 1 \right)}} } \right\}\left\{ {\hat{Y}_{{11}}^{{\left( 2 \right)}} ,\hat{Y}_{2}^{{\left( 2 \right)}} ,\hat{Y}_{8}^{{\left( 2 \right)}} , \ldots ,\hat{Y}_{8}^{{\left( 2 \right)}} } \right\},\left\{ {\hat{Y}_{1}^{{\left( p \right)}} ,\hat{Y}_{4}^{{\left( p \right)}} ,\hat{Y}_{4}^{{\left( p \right)}} , \ldots ,\hat{Y}_{{10}}^{{\left( p \right)}} } \right\}$$\end{document} Thus, a bootstrap sample of F values is formed. The critical value Fbootstrap is obtained by calculating percentages over this sample. The significance value is the probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P(Fbootstrap\:>\:Fvalue)$$\end{document}. Step 5. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P(Fbootstrap\:>\:Fvalue)<\alpha\:$$\end{document}, the DeepDent model is sufficient and proceeds to Step 6 for input significance tests. Otherwise, the DeepDent model is inadequate, and the algorithm is stopped. Step 6. All inputs are tested with the (19) hypotheses. The method in Step 4 is applied to calculate the critical value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Fbootstrap$$\end{document}. The significance value is calculated with the probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P(Fbootstrap\:>\:F\_value)$$\end{document}. Step 7. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P(Fbootstrap\:>\:F\_value)<\alpha\:$$\end{document}, the input is significant; otherwise, it is not significant. Step 8. Steps 6–7 are repeated for all inputs. An automatic forecasting method based on E-DeepDenT One of the most important problems in the application of artificial neural networks to the forecasting problem is the determination of hyperparameters and inputs. Practitioners find it difficult to access and understand the complex, detailed theoretical knowledge in forecasting methods. It is very important from a practical point of view that the selection of inputs and the determination of other hyperparameters in a forecasting method that can be used in practice is carried out with a set of test procedures instead of leaving it to the user. Fig. 1Flow-chart for Automatic Forecasting Method based on E-DeepDenT. Flow-chart for Automatic Forecasting Method based on E-DeepDenT. In this study, an automated forecasting algorithm based on the explainable DeepDenT method is created. The flow diagram of the algorithm is given in Fig. 1. The details of the automatic forecasting algorithm are presented in Algorithm 3. Algorithm 3: Explainable DeepDent automatic forecasting method based on an artificial neural network. Step 1. The lowest and highest values that the method can take for the hyperparameter and hyperparameters given below are determined. Minimum and Maximum Number of Hidden Layers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{min}$$\end{document} ve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{max}$$\end{document}.Number of minimum and maximum time steps: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{min}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{max}$$\end{document}.Number of repetitions for random starting weights: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}$$\end{document}Number of learning samples for training, validation, and test set: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train},{n}_{val}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{test}$$\end{document}.Period of the time series: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s$$\end{document}Number of bootstrap replications: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{bst}$$\end{document}Number of forecasts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}$$\end{document} Minimum and Maximum Number of Hidden Layers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{min}$$\end{document} ve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{max}$$\end{document}. Minimum and Maximum Number of Hidden Layers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{min}$$\end{document} ve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{max}$$\end{document}. Number of minimum and maximum time steps: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{min}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{max}$$\end{document}. Number of minimum and maximum time steps: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{min}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{max}$$\end{document}. Number of repetitions for random starting weights: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}$$\end{document} Number of repetitions for random starting weights: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}$$\end{document} Number of learning samples for training, validation, and test set: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train},{n}_{val}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{test}$$\end{document}. Number of learning samples for training, validation, and test set: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train},{n}_{val}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{test}$$\end{document}. Period of the time series: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s$$\end{document} Period of the time series: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s$$\end{document} Number of bootstrap replications: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{bst}$$\end{document} Number of bootstrap replications: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{bst}$$\end{document} Number of forecasts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}$$\end{document} Number of forecasts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}$$\end{document} Step 2. The time series is divided into three sets as training validation and test set as follows. The original time series \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{t}$$\end{document} consists of a total of n observations.22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{t}=\left\{{x}_{1},{x}_{2},\dots\:,{x}_{n}\right\}$$\end{document} The training data is the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{t}^{train}$$\end{document} consisting of the first \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train}$$\end{document} observations. Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train}1.645\sqrt{\frac{1+2(AC{F}_{1}+{\sum\:}_{i=2}^{m-1}AC{F}_{i}^{2})}{n}}$$\end{document} If Eq. (26) is satisfied, the time series contains seasonality, and seasonal differencing is applied by taking \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D=1$$\end{document} with Eq. (27). Otherwise, since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D=0$$\end{document}, no operation is applied.27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{t}=(1-{B}^{s}{)}^{D}{x}_{t}$$\end{document} In the second step, the presence of a unit root in the time series is investigated by applying the “Augmented Dickey-Fuller” test to the series, which is accepted to be seasonally stationary. If the time series contains a unit root, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d=1\:$$\end{document}is taken and differencing is applied with Eq. (28). Otherwise, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d=0$$\end{document} and no differencing is applied.28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}=(1-B{)}^{d}{z}_{t}$$\end{document} The “none” model is always used in the ADF test, and the best lag order is selected according to the AIC criterion. Step 4. Calculate the partial autocorrelation coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{kk},k=\mathrm{1,2},\dots\:,nlag$$\end{document} and the variances of the partial autocorrelations to calculate confidence intervals.29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{k}=\frac{\frac{1}{{n}_{trn}-k}\sum\:_{t=k+1}^{{n}_{trn}}\left({x}_{t}-\stackrel{-}{x}\right)\left({x}_{t-k}-\stackrel{-}{x}\right)}{\frac{1}{{n}_{trn}}\sum\:_{t=1}^{{n}_{trn}}{\left({x}_{t}-\stackrel{-}{x}\right)}^{2}}$$\end{document}30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{kk}=\frac{{r}_{k}-\sum\:_{j=1}^{k-1}{r}_{k-1,j}{r}_{k-j}}{1-{\sum\:}_{j=1}^{k-1}{r}_{k-1,j}{r}_{j}}$$\end{document}31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:V\left({r}_{kk}\right)=\frac{1}{{n}_{trn}}$$\end{document} Lags corresponding to the partial autocorrelation coefficients outside the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\pm\:2\sqrt{V\left({r}_{kk}\right)}\:$$\end{document}boundaries are determined. These lags form the elements of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:LagK$$\end{document}. The membership of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:LagK$$\end{document} is given by Eq. (32) with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k=\mathrm{1,2},...,nlag.$$\end{document}32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{LagK}\left(k\right)=\left\{\begin{array}{cc}1,&\:f\:\:{r}_{kk}>2\sqrt{V\left({r}_{kk}\right)}\:veya\:{r}_{kk}<-2\sqrt{V\left({r}_{kk}\right)}\\\:0,\:&\:otherwise\end{array}\right.$$\end{document} PACF can be used directly in determining linear AR models in theory. PACF does not guarantee the capture of non-linear AR structures. The aim here is to reduce the number of lagged variables to be used for the model. The test for non-linear relationships in the model can already be performed with input significance tests. Using PACF saves time for the algorithm but always carries the risk of missing some necessary NAR inputs. Step 5. The validation and test set performances are calculated by applying Steps 5.1 and 5.2 for all possible hyperparameter values. The possible hyperparameter sets are different combinations of elements taken from the sets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:q\in\:\left[{q}_{min},{q}_{max}\right]$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\in\:\left[{m}_{min},{m}_{max}\right]$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h\in\:\left[{h}_{min},{h}_{max}\right]$$\end{document}, totalling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{q}_{max}-{q}_{min}+1)\times\:({m}_{max}-{m}_{min}+1)\times\:({h}_{max}-{h}_{min}+1)$$\end{document}. Step 5.1. The set of random initial weights is randomly generated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}$$\end{document}times. Step 5.2. For each set of random initial weights, Step 5.2.1 and Step 5.2.3 are applied \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}\:$$\end{document}times. Step 5.2.1. The model validity and the significance of the inputs are tested by applying Algorithm 3 to the “DeepDenT” neural network with the current set of hyperparameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(m,q,h)$$\end{document} and lagged variables generated according to the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:LagK$$\end{document} set as input or Step 5.2.3. Step 5.2.2. If the model is not valid, calculate the validity and test set performances by taking \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:LagK=\left\{1\right\}$$\end{document} and proceed to Step 5.3. If the model is valid, proceed to Step 5.2.3. Step 5.2.3. Test the significance of each input in the model. If all inputs in the model are significant, calculate the validity and test set performances and proceed to Step 5.3. If some inputs are meaningless, remove the meaningless inputs from the model and return to Step 5.2.1 with the meaningful inputs. The purpose of the hypothesis test proposed here is to support the automatic forecasting method. In some cases, assumptions that arise for input significance tests can be disregarded. The goal can be seen as freeing the user from having to make their own choices. If the algorithm’s computation time is unreasonable, then the number of bootstrap repetitions can be sacrificed. In the automatic forecasting algorithm, there is a trade-off between computation time and the correct selection of the model. Step 5.3. Select the best one from the validity set results obtained for different random initial weight sets. Step 6. From the validity performances calculated for different hyperparameter sets, the best performance is selected. Step 7. Check the test performance for the selected hyperparameter set, calculate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}\:$$\end{document}forecasts from the trained model and terminate the algorithm. Simulation study for explainability robustness In the simulation study, the successful results of the proposed hypothesis tests for explainability were investigated. The simulation study was carried out on the time series generated from the 3 models given by Eqs. (33), (34) and (35).33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Model\:1:\:{x}_{t}=1+0.3{x}_{t-1}+0.4\sqrt{{x}_{t-2}}+{\epsilon\:}_{t}$$\end{document}34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Model\:2:\:{x}_{t}=1+0.3{x}_{t-1}+0.4\mathrm{l}\mathrm{o}\mathrm{g}\left({x}_{t-2}\right)+{\epsilon\:}_{t}$$\end{document}35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Model\:3:\:{x}_{t}=1+0.3\sqrt{{x}_{t-1}}+0.4\mathrm{l}\mathrm{o}\mathrm{g}\left({x}_{t-2}\right)+{\epsilon\:}_{t}$$\end{document} Initial values for the simulated time series are taken as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:0.3$$\end{document}. The residuals are normally distributed with a zero mean and a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:0.1$$\end{document} standard deviation. The length of the simulated time series is 500. The first 30 observations are discarded because of the burn-in period. I different parameter sets, 1000 simulated time series were generated. The simulated series was solved using DeepDent with four inputs: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x\left(t-1\right),\:x\left(t-2\right),\:x\left(t-3\right),\:x(t-4)$$\end{document}. The model is expected to be adequate if the first input x(t-1) is significant and the last three inputs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x(t-2),\:x(t-3),\:$$\end{document}and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:x(t-4)$$\end{document}are insignificant. The reason why \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x(t-2)$$\end{document} is expected to be insignificant is that, due to the time step in the deep neural network, lags greater than 1 can be automatically included in the model. The percentage of correct decisions for a type I error probability of 0.10 is given in Table 1 below. When Table 1 is examined, it is seen that a very high percentage of correct decisions was reached in the model validity tests. We also observe that the percentage of correct decisions in the model for the one-step lagged \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x\left(t-1\right)\:$$\end{document}variable is quite large. It is seen that the percentage of correct decisions in the variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x(t-2),\:x(t-3),\:$$\end{document}and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:x(t-4)$$\end{document} does not fall below 70. Table 1Simulation study results.ModelModel sufficiency testInput significance tests \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-2}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-4}$$\end{document} Model 10,94000,93200,72900,82100,8300Model 20,97200,94200,78100,79500,7690Model 30,96100,95000,74600,83100,8070 Table 1Simulation study results.ModelModel sufficiency testInput significance tests \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-2}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t-4}$$\end{document} Model 10,94000,93200,72900,82100,8300Model 20,97200,94200,78100,79500,7690Model 30,96100,95000,74600,83100,8070 Simulation study results. Applications to M3 and M4 data sets To investigate the performance of the proposed method, annual data from the M3 time series forecasting competition presented in29 and time series from the M4 time series competition presented in30 are used. The first results were obtained for the M4 time series. There are 100,000 time series in total in the M4 competition. This is a very large number for obtaining performance results. In this study, a total of 500 time series were used, 100 from each of the annual, monthly, quarterly, weekly, and hourly time series categories, so that the results can be obtained in a reasonable time on a personal computer. The first 100 time series of each category were selected to avoid any subjective selection and bias in the selection of time series. The benchmarks and the winning method in M4 are taken into account in the comparison of the results obtained. The results with and without the use of bootstrap hypothesis testing in the proposed automatic forecasting method are obtained separately. Table 2 presents the results obtained for the annual data in the M4 competition. In the table, two important error measures used in the competition are symmetric mean absolute percentage error (SMAPE) and mean absolute weighted error (MASE). The median of the values of the metrics for the 100 time series is shown in the table. This is because the experimental distribution of the 100-unit samples of MASE and SMAPE values is not symmetric. The results of the other methods in the M4 competition were obtained from the web, and similar calculations were made. The selection of hyperparameters in the application was made as shown below. Minimum and Maximum Number of Hidden Layers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{min=}1$$\end{document} ve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{max}=5$$\end{document}.Number of minimum and maximum time steps: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{min}=1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{max}=5$$\end{document}.Number of repetitions for random starting weights: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}=10$$\end{document}Number of learning samples for training, validation, and test set: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train},{n}_{val}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{test}$$\end{document} values are determined according to the observation number of time series with %10, %10 and %80 ratio.Number of bootstrap replications: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{bst}=1000$$\end{document}Number of forecasts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}$$\end{document} is determined according to the problem.cn: the number of chromosomes in a population: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cn=30$$\end{document}cr: crossover ratio: cr = 0.1.maxitr: the number of maximum iteration: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:maxitr=100$$\end{document}limit1: The limit value for the re-starting strategy: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:limit1=30$$\end{document}limit2: The limit value for the early stopping rule: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:limit2=50$$\end{document} Minimum and Maximum Number of Hidden Layers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{min=}1$$\end{document} ve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{max}=5$$\end{document}. Minimum and Maximum Number of Hidden Layers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{min=}1$$\end{document} ve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{m}_{max}=5$$\end{document}. Number of minimum and maximum time steps: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{min}=1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{max}=5$$\end{document}. Number of minimum and maximum time steps: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{min}=1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{max}=5$$\end{document}. Number of repetitions for random starting weights: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}=10$$\end{document} Number of repetitions for random starting weights: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{repeat}=10$$\end{document} Number of learning samples for training, validation, and test set: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train},{n}_{val}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{test}$$\end{document} values are determined according to the observation number of time series with %10, %10 and %80 ratio. Number of learning samples for training, validation, and test set: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{train},{n}_{val}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{test}$$\end{document} values are determined according to the observation number of time series with %10, %10 and %80 ratio. Number of bootstrap replications: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{bst}=1000$$\end{document} Number of bootstrap replications: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{bst}=1000$$\end{document} Number of forecasts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}$$\end{document} is determined according to the problem. Number of forecasts: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{f}$$\end{document} is determined according to the problem. cn: the number of chromosomes in a population: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cn=30$$\end{document} cn: the number of chromosomes in a population: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cn=30$$\end{document} cr: crossover ratio: cr = 0.1. cr: crossover ratio: cr = 0.1. maxitr: the number of maximum iteration: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:maxitr=100$$\end{document} maxitr: the number of maximum iteration: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:maxitr=100$$\end{document} limit1: The limit value for the re-starting strategy: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:limit1=30$$\end{document} limit1: The limit value for the re-starting strategy: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:limit1=30$$\end{document} limit2: The limit value for the early stopping rule: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:limit2=50$$\end{document} limit2: The limit value for the early stopping rule: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:limit2=50$$\end{document} Table 2Results obtained from analyzing the annual time series of the M4 competition.MethodMedian (SMAPE)Median (MASE)ARIMA15,16815,9229Com15,25216,0105Damped16,91415,8440ETS13,22484,6602Holt16,29516,0096MLP9,19024,4044Naive10,55444,4336Naive210,55444,4336RNN7,86463,5238SES10,55454,4220sNaive10,55444,4336Teta9,11323,9294Winner Method10,11144,431Auto-DeepDenT1,68613,1714Auto DeepDenT with Bootstrap2,99744,5080 Table 2Results obtained from analyzing the annual time series of the M4 competition.MethodMedian (SMAPE)Median (MASE)ARIMA15,16815,9229Com15,25216,0105Damped16,91415,8440ETS13,22484,6602Holt16,29516,0096MLP9,19024,4044Naive10,55444,4336Naive210,55444,4336RNN7,86463,5238SES10,55454,4220sNaive10,55444,4336Teta9,11323,9294Winner Method10,11144,431Auto-DeepDenT1,68613,1714Auto DeepDenT with Bootstrap2,99744,5080 Results obtained from analyzing the annual time series of the M4 competition. Table 3 presents the results obtained for the quarterly data. Table 3Results obtained from the analysis of the quarterly time series.MethodMedian (SMAPE)Median (MASE)ARIMA2,08120,9695Com2,69911,1049Damped2,75901,0935ETS2,24430,8323Holt2,53331,0335MLP5,11001,9226Naive4,67801,9154Naive24,67801,9154RNN5,14061,8390SES4,74931,9188sNaive4,6781,9154Teta3,46641,3583Winner Method2,56580,9316Auto-DeepDenT0,58450,8434Auto DeepDenT with Bootstrap0,57080,8896 Table 3Results obtained from the analysis of the quarterly time series.MethodMedian (SMAPE)Median (MASE)ARIMA2,08120,9695Com2,69911,1049Damped2,75901,0935ETS2,24430,8323Holt2,53331,0335MLP5,11001,9226Naive4,67801,9154Naive24,67801,9154RNN5,14061,8390SES4,74931,9188sNaive4,6781,9154Teta3,46641,3583Winner Method2,56580,9316Auto-DeepDenT0,58450,8434Auto DeepDenT with Bootstrap0,57080,8896 Results obtained from the analysis of the quarterly time series. Table 4 presents the results obtained for the monthly data. Table 4Results obtained in the analysis of monthly time series.MethodMedian (SMAPE)Median (MASE)ARIMA8,07620,7844Com7,91180,7499Damped6,93710,8553ETS6,83640,8475Holt9,18650,8360MLP11,29191,2932Naive9,01081,3649Naive27,24400,9998RNN11,18691,2892SES7,41530,9321sNaive8,04160,8607Teta8,17770,8217Winner Method7,05860,7156Auto-DeepDenT2,11270,8190Auto DeepDenT with Bootstrap2,07190,7715 Table 4Results obtained in the analysis of monthly time series.MethodMedian (SMAPE)Median (MASE)ARIMA8,07620,7844Com7,91180,7499Damped6,93710,8553ETS6,83640,8475Holt9,18650,8360MLP11,29191,2932Naive9,01081,3649Naive27,24400,9998RNN11,18691,2892SES7,41530,9321sNaive8,04160,8607Teta8,17770,8217Winner Method7,05860,7156Auto-DeepDenT2,11270,8190Auto DeepDenT with Bootstrap2,07190,7715 Results obtained in the analysis of monthly time series. Table 5 presents the results obtained for daily data. Table 5Results obtained in the analysis of the daily time series.MethodMedian (SMAPE)Median (MASE)ARIMA3,09864,0191Com2,67043,1697Damped2,94663,1620ETS2,98233,2142Holt3,20053,4203MLP5,96299,4773Naive2,78933,1084Naive22,78933,1084RNN3,55145,0533SES2,79293,1058sNaive2,78933,1084Teta2,66043,1663Winner Method2,29113,2913Auto-DeepDenT0,44592,7110Auto DeepDenT with Bootstrap0,44392,6941 Table 5Results obtained in the analysis of the daily time series.MethodMedian (SMAPE)Median (MASE)ARIMA3,09864,0191Com2,67043,1697Damped2,94663,1620ETS2,98233,2142Holt3,20053,4203MLP5,96299,4773Naive2,78933,1084Naive22,78933,1084RNN3,55145,0533SES2,79293,1058sNaive2,78933,1084Teta2,66043,1663Winner Method2,29113,2913Auto-DeepDenT0,44592,7110Auto DeepDenT with Bootstrap0,44392,6941 Results obtained in the analysis of the daily time series. Table 6 presents the results obtained for the hourly data. Table 6Results obtained in the analysis of the hourly time series.MethodMedian (SMAPE)Median (MASE)ARIMA7,50481,2718Com5,96671,1948Damped5,33781,1075ETS5,69351,1981Holt13,64262,1852MLP10,22671,7159Naive6,06551,2317Naive26,06551,2317RNN11,08441,7144SES6,16961,2820sNaive6,06551,2317Teta6,52191,3034Winner Method6,64291,121Auto-DeepDenT1,44890,9788Auto DeepDenT with Bootstrap1,36290,9620 Table 6Results obtained in the analysis of the hourly time series.MethodMedian (SMAPE)Median (MASE)ARIMA7,50481,2718Com5,96671,1948Damped5,33781,1075ETS5,69351,1981Holt13,64262,1852MLP10,22671,7159Naive6,06551,2317Naive26,06551,2317RNN11,08441,7144SES6,16961,2820sNaive6,06551,2317Teta6,52191,3034Winner Method6,64291,121Auto-DeepDenT1,44890,9788Auto DeepDenT with Bootstrap1,36290,9620 Results obtained in the analysis of the hourly time series. When the tables are analyzed, the proposed method can produce better results than the benchmarks in the competition and the winning method of the competition. Secondly, applications were made for 23,000 annual time series used in the M3 competition. In the application, the results are given based on the SMAPE criterion, as in the competition. However, instead of the mean statistic as in the competition, the median of the 23,000 SMAPE values found for each method was obtained as a more accurate statistic since this value would give biased results in a skewed distribution. Because of this fact, the rankings were given based on the median statistics. The results are presented in Table 7. Table 7Results obtained from analyzing the annual time series.MethodStatisticsSMAPEMethodStatisticsSMAPEAuto DeepDenT with BootstrapRanking1CombSHDRanking13Median0,0311Median0,1168RbfRanking3SmartFCSRanking17Median0,1074Median0,1183RD-ANNRanking5HoltRanking15Median0,1090Median0,1177Flors-Pearc1Ranking2WinterRanking16Median0,1072Median0,1177PPAutocastRanking4BjautoRanking14Median0,1083Median0,1170DampenRanking6ThetasmRanking18Median0,1095Median0,1221ThetaRanking8Flors-Pearc2Ranking21Median0,1125Median0,1255Robust TrendRanking9AutoANNRanking24Median0,1130Median0,1308ForcxRanking11Naive 2Ranking19Median0,1134Median0,1237ForecatsproRanking7Autobox3Ranking23Median0,1105Median0,1289Autobox2Ranking10SingleRanking20Median0,1131Median0,1244ARARMARanking12Autobox1Ranking22Median0,1135Median0,1275 Table 7Results obtained from analyzing the annual time series.MethodStatisticsSMAPEMethodStatisticsSMAPEAuto DeepDenT with BootstrapRanking1CombSHDRanking13Median0,0311Median0,1168RbfRanking3SmartFCSRanking17Median0,1074Median0,1183RD-ANNRanking5HoltRanking15Median0,1090Median0,1177Flors-Pearc1Ranking2WinterRanking16Median0,1072Median0,1177PPAutocastRanking4BjautoRanking14Median0,1083Median0,1170DampenRanking6ThetasmRanking18Median0,1095Median0,1221ThetaRanking8Flors-Pearc2Ranking21Median0,1125Median0,1255Robust TrendRanking9AutoANNRanking24Median0,1130Median0,1308ForcxRanking11Naive 2Ranking19Median0,1134Median0,1237ForecatsproRanking7Autobox3Ranking23Median0,1105Median0,1289Autobox2Ranking10SingleRanking20Median0,1131Median0,1244ARARMARanking12Autobox1Ranking22Median0,1135Median0,1275 Results obtained from analyzing the annual time series. Table 7 shows that the proposed method has a lower SMAPE value than all other methods in the competition. Conclusions and discussion In this study, the explainability studies of DeepDenT deep artificial neural network, and an automatic forecasting method based on the explainable DeepDenT neural network are presented. Especially in recent years, the applications of deep artificial neural networks in solving forecasting problems are frequently encountered in the literature. Unfortunately, artificial intelligence methods and artificial neural networks used in the forecasting problem do not have the following features that classical forecasting methods have. Testing the validity of the model used in the method. Testing the significance of the explanatory variables or inputs in the model used in the method, or their contribution to the model. Determining the effect of the coefficients in the model used in the method or the inputs of the model. Testing the validity of the model used in the method. Testing the validity of the model used in the method. Testing the significance of the explanatory variables or inputs in the model used in the method, or their contribution to the model. Testing the significance of the explanatory variables or inputs in the model used in the method, or their contribution to the model. Determining the effect of the coefficients in the model used in the method or the inputs of the model. Determining the effect of the coefficients in the model used in the method or the inputs of the model. All these details can be easily obtained in a linear model used for classical forecasting with good characteristics and appropriate data sets. A machine learning, artificial intelligence, or artificial neural network model, on the other hand, lacks these features and has a larger number of hyperparameters compared to the classical forecasting method. An objective method for determining the values of these hyperparameters has not yet been proposed to reduce the computation time. In addition to all these, in recurrent neural networks such as LSTM and GRU, which are the most widely used deep learning methods in the literature, the cells are realized as in the classical neuron model, as a summation of the incoming inputs. In addition, due to the fact that recurrent networks are trained with training algorithms based on derivative information, the problem of exploding and vanishing gradients arises. The LSTM neural network has found a solution to the exploding and vanishing gradient problem with gates, but these gates cause the number of parameters in the model to increase four times. Another important problem in solving the forecasting problem is the determination of lagged variables or model degrees. The proposed methods presented in this study provide solutions to these problems in the literature. The methods presented in this study provide solutions to these problems in the literature. The proposed DeepDenT artificial neural network can be explained as in the classical forecasting methods, Model validity tests. Input significance tests. Determining the effect of the coefficients in the model or the inputs of the model is possible. Model validity tests. Model validity tests. Input significance tests. Input significance tests. Determining the effect of the coefficients in the model or the inputs of the model is possible. Determining the effect of the coefficients in the model or the inputs of the model is possible. Bootstrap hypothesis testing, which does not require any distributional assumptions, was used for hypothesis testing. In this study, an automatic forecasting method based on explainable DeepDenT is presented. In the automatic forecasting method, hyper-parameter values are automatically selected by the user within predefined ranges, while lagged variables in the model or features of the model can be determined automatically. The proposed automatic forecasting method provides the practitioner with a forecasting method that does not require subjective decision-making. The proposed automatic forecasting method has been tested on many applications and time series to demonstrate its performance. It is observed that the proposed methods have a good performance on both stock market time series and M3 and M4 competition time series. The fact that comparisons were not made with more recent methods has been added to the study as a limitation of the study. We do not claim that the proposed method will be better than all methods in the literature for every data set. Although there is no problem in the computation time of the methods presented in the study in general, the computation time of the proposed automatic forecasting method will be high, especially in the case of obtaining forecasts for many products in a short period of time, such as retail sales forecasting. In addition, it is known that in problems such as retail sales forecasting, different products will affect each other’s sales figures, and it may be useful to deal with multivariate time series analysis. Therefore, converting the method into a multivariate version for such problems and increasing the computation time of the method by narrowing the ranges given for the hyperparameter seem to be the measures that can be taken. An effective method for reducing the computation time of the proposed method can be the estimation of hyperparameters. All these solutions and details give direction to future studies. There are different approaches to explainability in the literature. The linear model used in this approach provides the contribution of each input to the model, while linear models used for forecasting with partial dependence, the contribution and significance of each input to the model can be tested. However, the inability to perform this operation in the DeepDenT model is addressed in the explainability section of this study. A limitation of the study is that explainability cannot be ensured for each observation point.