Forecasting of customer demands for production planning by local -nearest neighbor models
Introduction
Accurate demand forecasts are essential for all companies in a supply chain since several decisions have to be made before the actual values of important variables are known. In addition to forecasts of resource and energy demands (Hahn et al., 2009, Weinert et al., 2011, Hong and Fan, 2016), electricity prices (Aggarwal et al., 2009, Weron, 2014), failure prognosis for maintenance planning (Peng et al., 2010, Sikorska et al., 2011, Fritzsche et al., 2014, Gao et al., 2015) or yield and quality predictions (Lieber et al., 2013, Colledani et al., 2014), forecasts of future customer demands are necessary. In particular, accurate demand forecasts are crucial in build to stock productions, for instance often occurring in the fast moving consumer goods industry. In this context, raw material orders as well as production programs base strongly on demand forecasts and customer demands are directly satisfied from finished goods inventories. Normally, forecasts are calculated on a weekly or monthly basis for a horizon of approximately one year (Fleischmann et al., 2015) based on time series of past customer orders (Tempelmeier, 2008). Due to various influencing factors, demand evolutions are often highly volatile. For example, the demand of a customer depends on the customer’s remaining product inventory, the number of available substitute products on the market, the success of marketing campaigns, the current economic, political, social or ecological situation as well as basic agreements (Wiendahl et al., 2007, Porter, 2008). Therefore, forecasting future customer demands can be a difficult task.
A typical demand planning process in a company consists of six steps: (i) data acquisition and preparation, (ii) statistical forecasting, (iii) judgmental forecasting, (iv) consensus forecasting, (v) release of the forecast as well as (vi) evaluation and monitoring (Kilger and Wagner, 2015, Petropoulos et al., 2016). In this context, the step of statistical forecasting is of particular importance since it provides forecasts by automatic algorithms that are used as a basis for the following steps. Therefore, the paper at hand focuses on statistical forecasting and neglects the other steps of demand planning. It has to be noted that this paper considers the forecasting of regular customer demand characterized by time series with integer values greater than zero. For applications regarding intermittent demand, which is characterized by a sporadic demand evolution including periods without a demand, such as often observed for spare parts, see (Syntetos and Boylan, 2001, Wallström and Segerstedt, 2010, Kourentzes, 2013, Kourentzes, 2014, Van Wingerden et al., 2014).
Frequently, manufacturing companies have to forecast customer demands for thousands of products simultaneously. Hence, automatic forecasting methods have to be applied to calculate big amounts of forecasts. While simple forecasting methods, such as moving averages, simple exponential smoothing or linear regression methods are still often used in industrial practice (Küsters et al., 2006), more sophisticated methods already proved applicability in many practical applications as well as in forecasting competitions (Hyndman, 2020). Autoregressive integrated moving average (ARIMA) models (Hyndman and Khandakar, 2008, Babai et al., 2013, Box et al., 2015, Svetunkov and Boylan, 2019) and complex variants of exponential smoothing (Gardner, 2006, Hyndman et al., 2008, Ferbar Tratar et al., 2016, Kück et al., 2016b, Sbrana and Silvestrini, 2014, Sbrana and Silvestrini, 2019) are established methods that have shown potential in several studies, such as the M3-Competition (Makridakis and Hibon, 2000). In addition, machine learning methods, such as artificial neural networks (Crone et al., 2008, Crone and Kourentzes, 2010, Adeodato et al., 2011, Montavon et al., 2012, Kourentzes, 2013, Kourentzes et al., 2014) or support-vector machines (Cortes and Vapnik, 1995, Crone et al., 2008, Lu, 2014) have shown promising results in current studies, such as the NN3-Competition (Crone et al., 2011). However, these methods generally need long computation times and they carry the risk of overfitting when they are applied fully automatically. Moreover, some of these methods need long time series of past customer demands in order to fit their parameters in a suitable way. While demand forecasting can commonly be considered as a big data problem of calculating many forecasts for several products at regular intervals, it is nevertheless often the case that the individual demand histories only comprise a few years. Hence, computationally efficient methods that achieve a stable forecasting performance and do not need long demand histories to fit their parameters suitably would be beneficial.
Despite the variety of methods that have been applied to forecast industrial time series, local -nearest neighbor models have been mostly neglected so far. Based on a modeling by nonlinear dynamical systems, these methods are able to reconstruct the dynamical properties of a demand evolution as well as influencing factors out of a scalar time series of past customer demands. These so-called methods of nonlinear dynamics have been successfully applied to conduct different tasks in production and logistics systems, such as modeling and control of production systems (Scholz-Reiter et al., 2002, Papakostas et al., 2009) or forecasting sporadic spare parts demand (Nikolopoulos et al., 2016). Furthermore, small-scale studies covering one-step ahead forecasts with small sets of exemplary customer demand time series have already shown promising results of local -nearest neighbor forecasting models of nonlinear dynamics to forecast regular monthly customer demand (Mulhern and Caprara, 1994, Kück and Scholz-Reiter, 2013, Kück et al., 2014). However, the exemplary results of these studies can only be seen as an indication that these methods could perform well to forecast industrial time series and this assumption has to be proven in a comprehensive study.
The contributions of this paper are the following:
- •
Proposition of local -nearest neighbor models for regular monthly demand forecasting: The paper proposes local -nearest neighbor models as a novel class of methods to forecast regular monthly customer demands of a manufacturing company.
- •
Parameter comparison and model selection strategies: Two different versions of -nearest neighbor models are compared, namely locally constant models and locally linear regression models. While standard parameters are used for delay coordinate embedding, different regularization methods with respective regularization parameters are compared regarding their suitability to forecast industrial time series. Moreover, different model selection strategies are proposed.
- •
Evaluation regarding forecast accuracy: The forecasting performance of two pre-parameterized versions of the local -nearest neighbor models as well as different individual model selection strategies are evaluated in a comprehensive empirical study utilizing 808 monthly industrial time series of the M3-Competition. As a first criterion of forecasting performance, the methods are evaluated in rolling-origin experiments with forecasts from multiple origins across short, medium and long horizons in comparison to several established benchmark methods. In addition to the evaluations in terms of mean forecast errors, Friedman and Nemenyi tests are conducted to assess significant differences between the performances of the local -nearest neighbor models and the benchmark methods.
- •
Evaluation regarding computation times: In contrast to the usual practice in other studies, this paper also evaluates the computation times of the forecasting methods. In this context, mean times per parameter optimization and per forecast are compared.
- •
Evaluation regarding inventory performance: Another criterion that is often neglected in other studies is the inventory performance achieved based on calculated forecasts. This paper evaluates the local -nearest neighbor models and the benchmark methods considering achieved service levels and needed safety stocks in an inventory simulation.
The remainder of this paper is structured as follows: Section 2 provides an overview of the existing literature regarding applications of local -nearest neighbor forecasting models. Section 3 explains the theoretical background. After a parameter comparison and a comparison of different model selection strategies in Section 4, Section 5 comprises a comprehensive empirical study to assess the forecasting performance of the local -nearest neighbor models in terms of forecast errors and computation times. Section 6 applies the forecasting methods in an inventory simulation and evaluates their performance regarding inventory metrics. Section 7 concludes the paper.
Section snippets
Literature overview
Local -nearest neighbor models belong to the methods of nonlinear dynamics, which originate from the theory of dynamical systems and which base on the paradigm of deterministic chaos. For general overviews regarding methods of nonlinear dynamics, see Kantz and Schreiber, 2004, Aguirre and Letellier, 2009 and Broer and Takens (2011). Apart from successful applications in modeling, analysis and control of manufacturing systems (Ramadge, 1993, Prabhu and Duffie, 1995, Wiendahl and Scheffczyk, 1999
Theoretical background
This section describes the necessary steps to build a local -nearest neighbor (nn) forecasting model. After an introductory overview, the steps of delay coordinate embedding, nn searching and building a local nn forecasting model are detailed.
Parameter comparison and model selection strategies
This section determines valuable parameter configurations and model selection strategies for the local nn models to forecast monthly customer demands. The section starts with an experiment description, considering the compared parameter configurations, the used dataset and the performance criteria, which are mean forecast accuracies and computation times. Afterwards, the results of the parameter comparison are shown and suitable model selection strategies are determined.
Empirical study 1: Forecast errors and computation times
This section evaluates the performance of the four model selection strategies for local nn forecasting models determined in the last section. After an experimental description, the section comprises evaluations regarding forecast errors as well as times per forecast and per optimization.
Empirical study 2: Inventory simulation
After evaluating the performance of the forecasting methods in terms of forecast accuracy and computation times, this section evaluates their performance in inventory simulations. For this purpose, a discrete-event simulation model has been implemented. For a given time series of customer demands, each of the forecasting methods is applied to forecast the demands and a periodic order-up-to policy is used to calculate product replenishment orders. Section 6.1 describes the simulation model and
Conclusion
This paper has assessed the suitability of local -nearest neighbor models to forecast industrial time series data. After a thorough description of the theoretical background, an extensive parameter comparison has been conducted to find reasonable parameters and model selection strategies that lead to stable forecasting performance. In this context, the locally constant nn model was tested with a local averaging according to an arithmetic mean and a median. Moreover, different regularization
CRediT authorship contribution statement
Mirko Kück: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing - original draft, Writing - review & editing, Visualization. Michael Freitag: Conceptualization, Writing - review & editing, Supervision, Project administration.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (140)
- et al.
MLP ensembles improve long term prediction accuracy over single networks
Int. J. Forecast.
(2011) - et al.
Electricity price forecasting in deregulated markets: A review and evaluation
Int. J. Electr. Power Energy Syst.
(2009) - et al.
Forecasting and inventory performance in a two-stage supply chain with ARIMA(0,1,1) demand: Theory and empirical analysis
Int. J. Prod. Econ.
(2013) - et al.
A review and comparison of strategies for multi-step ahead time series forecasting based on the NN5 forecasting competition
Expert Syst. Appl.
(2012) - et al.
Conditionally dependent strategies for multiple-step-ahead prediction in local learning
Int. J. Forecast.
(2011) - et al.
Nonlinear deterministic forecasting of daily dollar exchange rates
Int. J. Forecast.
(1999) Nonlinear prediction of chaotic time series
Physica D
(1989)- et al.
Distinguishing between stochastic and deterministic behavior in high frequency foreign exchange rate returns: Can non-linear dynamics help forecasting?
Int. J. Forecast.
(1996) - et al.
Design and management of manufacturing systems for production quality
CIRP Ann. Manuf. Technol.
(2014) - et al.
Advances in forecasting with neural networks? Empirical evidence from the NN3 competition on time series prediction
Int. J. Forecast.
(2011)
Feature selection for time series prediction - A combined filter and wrapper approach for neural networks
Neurocomputing
Nonlinear characterization of the performance of production and logistics networks
J. Manuf. Syst.
Demand forecasting with four-parameter exponential smoothing
Int. J. Prod. Econ.
Exchange-rate forecasts with simultaneous nearest-neighbour methods: evidence from the EMS
Int. J. Forecast.
Dynamics of resource sharing in production networks
CIRP Ann. Manuf. Technol.
Optimal prognostic distance to minimize total maintenance cost: The case of the airline industry
Int. J. Prod. Econ.
Cloud-enabled prognosis for manufacturing
CIRP Ann. Manuf. Technol.
Exponential smoothing: The state of the art-Part II
Int. J. Forecast.
Electric load forecasting methods: Tools for decision making
European J. Oper. Res.
Probabilistic electric load forecasting: A tutorial review
Int. J. Forecast.
Understanding supply chain dynamics: A chaos perspective
European J. Oper. Res.
A brief history of forecasting competitions
Int. J. Forecast.
Neighbourhood selection for local modelling and prediction of hydrological time series
J. Hydrol.
Short-term wind power forecasting using evolutionary algorithms for the automated specification of artificial intelligence models
Int. J. Forecast.
Markov Chain model for turbulent wind speed data
Physica A
Intermittent demand forecasts with neural networks
Int. J. Prod. Econ.
On intermittent demand model optimisation and selection
Int. J. Prod. Econ.
Neural network ensemble operators for time series forecasting
Expert Syst. Appl.
Emergence of non-predictable dynamics caused by shared resources in production networks
Procedia CIRP
Robust methods for the prediction of customer demands based on nonlinear dynamical systems
Procedia CIRP
Regularized local linear prediction of chaotic time series
Physica D
Forecasting software: Past, present and future
Int. J. Forecast.
A semi-parametric approach for estimating critical fractiles under autocorrelated demand
European J. Oper. Res.
Quality prediction in interlinked manufacturing processes based on supervised & unsupervised machine learning
Procedia CIRP
Is a random walk the best exchange rate predictor?
Int. J. Forecast.
A comparison between neural networks and chaotic models for exchange rate prediction
Comput. Statist. Data Anal.
Sales forecasting of computer products based on variable selection scheme and support vector regression
Neurocomputing
The M3-competition: Results, conclusions and implications
Int. J. Forecast.
Local averaging optimization for chaotic time series prediction
Neurocomputing
A comparison of the accuracy of short term foreign exchange forecasting methods
Int. J. Forecast.
Multiple-supplier inventory models in supply chain management: A review
Int. J. Prod. Econ.
A nearest neighbor model for forecasting market response
Int. J. Forecast.
Forecasting supply chain sporadic demand with nearest neighbor approaches
Int. J. Prod. Econ.
Forecasting with cue information: A comparison of multiple regression with alternative forecasting approaches
European J. Oper. Res.
Modelling the complexity of manufacturing systems using nonlinear dynamics approaches
CIRP Ann. Manuf. Technol.
Do big losses in judgmental adjustments to statistical forecasts affect experts behaviour?
European J. Oper. Res.
The inventory performance of forecasting methods: Evidence from the M3 competition data
Int. J. Forecast.
Modelling and analysis of nonlinear dynamics in autonomous heterarchical manufacturing systems control
CIRP Ann. Manuf. Technol.
Prediction in chaotic nonlinear systems: Methods for time series with broadband Fourier spectra
Phys. Rev. A
Modeling nonlinear dynamics and chaos: A review
Math. Probl. Eng.
Cited by (56)
On the value of operational flexibility in the trailer shipment and assignment problem: Data-driven approaches and reinforcement learning
2023, International Journal of Production EconomicsA Lie group kernel learning method for medical image classification
2023, Pattern RecognitionIntegrated lot-sizing and scheduling: Mitigation of uncertainty in demand and processing time by machine learning
2023, Engineering Applications of Artificial IntelligenceCitation Excerpt :We consider three factors influencing the demand as follows: Generally, time-varying features are widely used as well-known inputs in the prediction of the demand Kück and Freitag (2021). Therefore, we consider the dependency between the amount of demand and the period’s number (e.g., seasonal changes) as the first factor.
Decision support framework for inventory management combining fuzzy multicriteria methods, genetic algorithm, and artificial neural network
2022, Computers and Industrial EngineeringReview of load forecasting based on artificial intelligence methodologies, models, and challenges <sup>✰</sup>
2022, Electric Power Systems Research