The deeper knowledge in the field of study was comprehensive to determine the specialized area of my future profession. This gives me further motivation and passion to dedicate myself to keep moving further and deeper in machine learning/deep learning. Third, I would like to express my sincere thanks to the Korean government for a grant from the National Research Foundation of Korea (NRF) funded (MSIT; NRF-2019R1A2C4070310) that funded this work to greatly contribute to the field of study.
This is a critical turning point in life to equip state-of-the-art skills that qualify as a competent scholar. A sparse set of external variables (preliminary demand order and the number of non-working days) were considered in the NARX model. The performance evaluation resulted in the root mean square error (RMSE) of the proposed model being better than an ARIMA model in both training (18%) and testing (15%) sets.
The promising results of the proposed NARX model can be crucial for improving manufacturing planning to effectively reduce carrying costs and prevent inventory.
Problem Statement
The bullwhip effect can be caused by human actions, such as misunderstandings regarding supply and demand facts. Request forecast updating, order grouping, market volatility, rationing, and scarcity cooperation have all been established as causes of the bullwhip effect. The bullwhip effect is created by both of these forces working together, as well as the chain's technology and order managers' sound decision-making.
Since the extent of buffer stock plays a role in the bullwhip effect, it stands to reason that while the turnaround times between product replenishments are coupled with a longer supply chain; the fluctuation would be much greater (Lee, Padmanabhan, & Whang, 1997). As a result, it is verified that the impact of the bullwhip resulted in poorly coordinated local demand forecasting. The magnitude of the bullwhip effect is influenced by the responsiveness of the forecast to demand, an extension of this finding.
The bullwhip effect is increased by forecasts that are highly sensitive to changes in demand, while it is reduced by forecasts that are less responsive (Barlas and Gunduz, 2011).
Exiting Studies
If each stage uses local uncoordinated forecasts focused on incoming orders or demand, the bullwhip effect, or duplication of orders in the supply chain, occurs in all situations and policies. In addition, a study of ARIMA was conducted in the same area of the plastics industry to dive deeper to check the improvement of performance by comparing with other methods (Udom and Phumchusri, 2014). Forecasting a ten-month future period of demand using IBM SPSS application to vary the parameter resulted in the best fit in ARIMA(1, 0, 1) with 12% SE.
Since forecasting using ANN resulted in better error values and was suitable for both short-term and long-term demand forecasting (Al-Saba and El-Amin, 1999), ANN is also considered an accurate model in the auto parts manufacturing company. The aim of the study was to determine the best training approach for forecasting demand signals in the supply chain using the ANN method. Moreover, ANN has been applied in the field of renewable energy to forecast the electricity demand of smart grid technology.
Also, in the field of water supply, HLSALOA gave the highest degree of accuracy, according to the results of the applied ANN algorithms.
Methodology
However, increasing the number of input parameters relevant to the loads, such as general human activities in the field, temperature, humidity and other meteorological parameters, is expected to improve forecast accuracy (Akarslan and Hocaoglu, 2018). The Mahalanobis distance was used to find outliers in the multivariate sample, and the Granger test was used to decide which variables could be included in the model. Finally, the NARX model was compared with the univariate NAR and the persistence model to assess the contribution of the variables used in the NARX model to the prediction.
22 outliers were identified in the raw dataset, removed and replaced using the one-point linear tendency. The Granger test in La Mata, Oaxaca, revealed that solar radiation is the only element that affects wind speed production, making its configuration as an input vector in the NARX model crucial. However, very limited studies have been conducted to develop the NARX method to predict the demand of a first-tier supplier in the automotive industry, even though several external factors may influence the actual demand.
In this sense, a study is needed to develop a NARX model for demand forecasting considering the characteristics of the first-tier supplier in the automotive industry.
Purpose of Study
Data Sets
Parsimonious Variables Selection
Two key variables (pre-order amount and number of holidays) among the three external variables were selected to avoid multicollinearity problems in the prediction model. High correlation between extraneous variables can lead to multicollinearity problems that can degrade the accuracy of the prediction model. First, we performed a stepwise regression analysis, which showed that the number of holidays (p < 0.001) and the amount of the pre-order (p < 0.001) were included in the model; however, the volume of sales declined.
In addition, this study analyzed the correlation between the variables and found that sales volume is highly correlated with pre-order quantity (0.813 to 0.780) and number of holidays (0.734 to 0.725). This means that the amount of pre-order and the number of holidays without sales volume can correctly predict demand, as they have a high correlation with sales volume.
Time-lag Analysis
The number of input-output lags for the network was determined by autocorrelation function (ACF) and partial autocorrelation function (PACF) (Box et al., 1994). Based on the Figures 12 and 13, the coefficients of AFC gradually decrease to 0 as "tails off", and PACF shows that, after the high correlation in lag 1 (0.96) and the other's correlation sharply decreases as "cut off" below and close to the confidence interval.
- Autoregressive
- Moving Average
- Autoregressive Integrated Moving Average
- Non-seasonal ARIMA
- Seasonal ARIMA
- Time Series Using Shallow Networks
- Proposing Nonlinear Autoregressive with External (Exogenous) Input (NARX)
A variation of the error term εt would only affect the magnitude of the sequence, but not the trends. To determine the moving average, it is necessary to determine the number and value q of the parameters of the moving average from w1 to wq, taking into account some restrictions on the values, so that the process is stationary. The moving average model works well for stationary results, which are a type of time series that has no pattern or seasonality.
In a regression-like algorithm, a moving average model uses historical forecast errors instead of previous values of the forecast component. The variance of the error term εt, like that of autoregressive models, can only change the magnitude of the sequence, not the trends (Hyndman and Athanasopoulos, 2018). To determine the formula, the analyst first selects the required model, including the number of terms p, d, and q; then, using a nonlinear least squares procedure, determines the coefficients and provides a refined suggestion of the model parameters (Hyndman, and Athanasopoulos, 2018).
A non-seasonal ARIMA model is created using a combination of autoregressive differencing and a moving average model. The scale of the prediction intervals is also affected by the value of d; the larger the value of d, the larger the prediction intervals become. When d = 0, the standard deviation of the long-term forecast is equal to the standard deviation of the historical results, resulting in exactly the same forecast intervals (Hyndman and Athanasopoulos, 2018).
The seasonal portion of the model consists of concepts identical to the non-seasonal elements, but which include seasonal shifts. By changing the values of the correlations (weights) between components, a neural network can be trained to perform a specific role. Based on a performance and target calculation, the network is balanced so that the network output meets the target.
It can also be used for non-linear filtering with a noise-free variant of the input signal as the target output. Another essential use of the NARX network would be in the simulation of nonlinear dynamical systems. Past values of the output signal and past values of an isolated (exogenous) input signal are used to regress the next value of a target variable output signal yt.
As part of the traditional NARX architecture, such as the parallel architecture, the output data is fed back to the feed-forward neural network data.
- Loss and Accuracy Comparison
- Predicted Value Comparison
- Discussion
- Conclusion
Furthermore, the node number for the hidden layer was determined as 4 by analyzing the learning performance for different node numbers of the network. Approaching greater accuracy in forecasting actual demand placed by the automotive industry prime contractor aims to alleviate uncertainty in inventory management. The supplier tends to eliminate waste and unnecessary processes to maintain safety stock that minimizes costs in maintenance and expiration date of the individual product, while the high stock level can guarantee to satisfy the order.
In addition, the forecasts also affect the dilemma of the supply chain system in positioning factories and warehouses. Since the model is built based on a specific dataset in the given scenario of a first-line supplier, adapting this research result in other companies, parameters will have to change based on the different complexity and characteristics of the respective dataset. Impact of the length of the observed records on the performance of ANN and of conceptual parsimonious models for predicting storm runoff.
Demand forecasting and sharing strategies to mitigate fluctuations and the bullwhip effect in supply chains. A nonlinear autoregressive exogenous (NARX) neural network model for the prediction of daily direct solar radiation. Modeling and forecasting of flood water levels using the NARX neural network: case study at the Kelang River.
Pros and cons of using artificial neural networks versus logistic regression to predict medical outcomes. A comparative study between time series model and ARIMA model for sales forecasting for distributor in plastics industry. Sustainability in multi-tier supply chains: Understanding the dual agency role of the first-tier supplier.
