Machine Learning Demand Forecasting Based on ARIMA and Support Vector Machines to Improve Supply Chain Efficiency

Predictive analytics has become a crucial tool in modern business decision-making, particularly in demand forecasting. By leveraging historical data and statistical modeling techniques, businesses can enhance resource allocation, optimize inventory management, and improve overall operational efficiency. Among various predictive analytics methods, time series models are widely employed due to their ability to capture trends, seasonality, and patterns in demand fluctuations. This explores the application of time series models, including Moving Average (MA), Autoregressive (AR), Autoregressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA), Exponential Smoothing (Holt-Winters method), and advanced deep learning models such as Long Short-Term Memory (LSTM) networks. These models enable businesses to make data-driven decisions by accurately predicting future demand, thereby reducing costs and improving supply chain responsiveness. Furthermore, the integration of predictive analytics in business resource allocation extends beyond inventory management. It influences workforce planning, financial forecasting, and supply chain optimization, ensuring that organizations align their operations with anticipated market trends. However, despite the benefits, challenges such as data quality issues, model interpretability, external disruptions, and scalability concerns remain key obstacles in implementing time series forecasting. Emerging advancements in artificial intelligence (AI), big data analytics, and cloud-based forecasting solutions are shaping the future of predictive analytics. The integration of real-time data and machine learning models offers new opportunities for businesses to enhance forecasting accuracy and agility. As demand forecasting continues to evolve, organizations must adopt a strategic approach to model selection and implementation to fully capitalize on the advantages of predictive analytics. This study provides a comprehensive overview of predictive analytics for demand forecasting, highlighting the role of time series models in optimizing business resource allocation and discussing future trends that will drive further innovation in this field.

Multi-echelon supply chain networks face increasing complexity in demand forecasting and inventory optimization due to volatile market conditions and dynamic customer preferences. Traditional forecasting methods often struggle to capture long-range dependencies and complex temporal patterns inherent in supply chain data. Transformer-based architectures, originally developed for natural language processing (NLP), have emerged as powerful tools for time series forecasting in supply chain management (SCM). These models leverage self-attention mechanisms to process sequential data and capture intricate relationships across multiple time steps. This review examines the application of transformer models in demand forecasting and inventory optimization within multi-echelon supply chain networks. The paper analyzes how transformer architectures address challenges such as bullwhip effect amplification, demand volatility, and coordination across supply chain tiers. Key findings indicate that transformer-based approaches outperform conventional methods including autoregressive integrated moving average (ARIMA), long short-term memory (LSTM) networks, and traditional machine learning (ML) algorithms in forecast accuracy and computational efficiency. The review synthesizes recent developments in transformer architectures specifically adapted for supply chain contexts, including modifications for handling sparse data, incorporating external factors, and enabling real-time decision support. Furthermore, the paper explores integration strategies between demand forecasting and inventory optimization, examining how transformer predictions inform safety stock calculations, reorder point determination, and dynamic replenishment policies. Emerging trends such as attention mechanism interpretability, federated learning (FL) for collaborative forecasting, and hybrid models combining transformers with reinforcement learning (RL) are discussed. The review identifies critical gaps in current research, including limited validation in real-world multi-echelon settings, computational scalability challenges, and the need for robust frameworks addressing demand uncertainty propagation across supply chain levels.

The Covid-19 pandemic has caused severe economic depression and has disrupted the supply chains of various industries. The automobile industry which contributes significantly to the Indian economy was gravely hit due to the lockdowns, semiconductor shortage and the uncertainty associated with the pandemic. This research paper analyses the effect of Covid-19 on the automobile sales in India using the time series modelling approach. The data recorded by SIAM from 2012 to 2019 was used to develop the Autoregressive Integrated Moving Average (ARIMA) model following the Box-Jenkins methodology. ARIMA model (2, 1, 3) was chosen as it had the lowest AIC and BIC criteria. This model was used to forecast the sales from 2020 to 2021 to give a picture of the expected automobile sales had the pandemic not occurred. The forecasted data from the model developed has then been compared with actual automobile sales data during the pandemic to gauge the level of impact Covid-19 had on the Indian automobile industry. The paper also explores the associated challenges that the automobile industry had to face due to the pandemic.

Time series analysis of aerosol optical depth over New Delhi using Box–Jenkins ARIMA modeling approach

In response to the requirements of the green and low-carbon demands in the electricity supply chain, it is crucial to propose effective demand forecasting models to balance supply and demand of electrical materials and enhance operational efficiency. This paper introduces a combined forecasting model Seasonal Autoregressive Integrated Moving Average (SARIMA)-BP. The SARIMA-BP hybrid forecasting model enhances prediction accuracy by leveraging Autoregressive Integrated Moving Average (ARIMA) for initial forecasting and using its results as input features for BP neural networks (BPNN) to further refine the forecasts. Using various electrical materials in power projects as examples, the model verifies the accuracy of demand forecasts by comparing predicted quantities with actual demands. Furthermore, the model's performance superiority is ensured through comparison with results from other forecasting models. In the demand forecasting for steel-core aluminum stranded wire, the Mean Absolute Percentage Error (MAPE) value of the combined forecasting model is 9.8%, surpassing that of the other two models. This integrated approach not only meets the intricate forecasting demands of the electricity industry but also aids in optimizing resource allocation and improving operational efficiency, thereby reducing costs and achieving low-carbon sustainable development in the electricity sector.

A single ARIMA model cannot meet higher standards of prediction accuracy. Moreover, it can only deal with small prediction periods in the forecasting work. For the sake of prediction accuracy, we combined an ARIMA model with BP neural network. Firstly, an ARIMA forecasting model is established. Secondly the BP neural network is used to improve the single ARIMA model. The residual of ARIMA model is trained and fitted by BP neural network. Finally, more accurate results are given through combination with the forecast results of ARIMA model. The practice turns out that, compared with single ARIMA model, the prediction accuracy of new ARIMA model improved by BP neural networks is obviously enhanced, with an average error of forecast decreasing 10.4% by a large margin. Thus, the combined model proposed by this paper can be used in future prediction researches and industrial data analysis.

ARIMA forecasting of primary energy demand by fuel in Turkey

On March 16, 2020, in response to the emergence of the COVID-19 virus, the United States implemented an indefinite quarantine mandate. This unprecedented measure had repercussions, affecting businesses, schools, and government facilities across the nation. The mandate required these entities to either operate at reduced capacity or to halt operations altogether. Consequently, the stock market experienced extreme volatility during this period. This study aims to conduct an in-depth analysis of the responses to the pandemic of two firms operating in different industries. Amazon, an e-commerce and technology giant in the broadband retail industry, found itself with unprecedented demand for online shopping and cloud services considering the prevalent supply chain disruptions at the time. On the flip side, Darden Restaurants, a prominent player in the restaurant industry, had to either operate at a limited capacity, or shut down most of its restaurants altogether in coordination with the mandated closures issued by the government. This study will employ time series analysis techniques, specifically Autoregressive (AR), Moving Average (MA), and ARIMA (Autoregressive Integrated Moving Average) models. The objective is to assess the influence of COVID-19 on the selected firms and extract valuable insights from the data.

To identify the stochastic autoregressive integrated moving average (ARIMA) model for short term forecasting of hepatitis C virus (HCV) seropositivity among volunteer blood donors in Karachi, Pakistan. Ninety-six months (1998-2005) data on HCV seropositive cases (1000(-1) x month(-1)) among male volunteer blood donors tested at four major blood banks in Karachi, Pakistan were subjected to ARIMA modeling. Subsequently, a fitted ARIMA model was used to forecast HCV seropositive donors for 91-96 mo to contrast with observed series of the same months. To assess the forecast accuracy, the mean absolute error rate (%) between the observed and predicted HCV seroprevalence was calculated. Finally, a fitted ARIMA model was used for short-term forecasts beyond the observed series. The goodness-of-fit test of the optimum ARIMA (2,1,7) model showed non-significant autocorrelations in the residuals of the model. The forecasts by ARIMA for 91-96 mo closely followed the pattern of observed series for the same months, with mean monthly absolute forecast errors (%) over 6 mo of 6.5%. The short-term forecasts beyond the observed series adequately captured the pattern in the data and showed increasing tendency of HCV seropositivity with a mean +/- SD HCV seroprevalence (1000(-1) x month(-1)) of 24.3 +/- 1.4 over the forecast interval. To curtail HCV spread, public health authorities need to educate communities and health care providers about HCV transmission routes based on known HCV epidemiology in Pakistan and its neighboring countries. Future research may focus on factors associated with hyperendemic levels of HCV infection.

Because of its critical position in open economies and its extremely high volatility, the stock market price index has been a popular subject of market research. In modern financial markets, traders and practitioners have had trouble predicting the stock market price index. In order to solve this problem, some methods have been researched by researchers and suitable methods have been found. To analyze and forecast monthly stock market price index, a variety of statistical and econometric models are extensively used. Thus, this study aims to investigate the application of autoregressive integrated moving averages (ARIMA) for forecasting monthly stock market price index in Istanbul for the period from 2009- M01 to 2021-M03. As compared to all other tentative models, the research showed that the ARIMA (3,1,5) model is the best fit model for predicting the stock market price index. Forecasting is conducted by using the developed model ARIMA (3,1,5) and the results indicated that the forecasted values are very similar to the actual ones, reducing forecast errors. In general, the stock market price index in Istanbul; showed a downwards trend over the forecasted period. The results of the study can set an example for researchers and practitioners working in the stock market and can be a guide for economic decision units and investors in the stock market. 

Consumption of natural gas, a major clean energy source, increases as energy demand increases. We studied specifically the Turkish natural gas market. Turkey’s natural gas consumption increased as well in parallel with the world‘s over the last decade. This consumption growth in Turkey has led to the formation of a market structure for the natural gas industry. This significant increase requires additional investments since a rise in consumption capacity is expected. One of the reasons for the consumption increase is the user-based natural gas consumption influence. This effect yields imbalances in demand forecasts and if the error rates are out of bounds, penalties may occur. In this paper, three univariate statistical methods, which have not been previously investigated for mid-term year-ahead monthly natural gas forecasting, are used to forecast natural gas demand in Turkey’s Sakarya province. Residential and low-consumption commercial data is used, which may contain seasonality. The goal of this paper is minimizing more or less gas tractions on mid-term consumption while improving the accuracy of demand forecasting. In forecasting models, seasonality and single variable impacts reinforce forecasts. This paper studies time series decomposition, Holt-Winters exponential smoothing and autoregressive integrated moving average (ARIMA) methods. Here, 2011–2014 monthly data were prepared and divided into two series. The first series is 2011–2013 monthly data used for finding seasonal effects and model requirements. The second series is 2014 monthly data used for forecasting. For the ARIMA method, a stationary series was prepared and transformation process prior to forecasting was done. Forecasting results confirmed that as the computation complexity of the model increases, forecasting accuracy increases with lower error rates. Also, forecasting errors and the coefficients of determination values give more consistent results. Consequently, when there is only consumption data in hand, all methods provide satisfying results and the differences between each method is very low. If a statistical software tool is not used, time series decomposition, the most primitive method, or Winters exponential smoothing requiring little mathematical knowledge for natural gas demand forecasting can be used with spreadsheet software. A statistical software tool containing ARIMA will obtain the best results.

Background: Transportation demand forecasting is an essential activity for logistics operators and carriers. It leverages business operation decisions, infrastructure, management, and resource planning activities. Since 2015, there has been an increase in the use of deep learning models in this domain. However, there is a gap in works comparing traditional statistics and deep learning models for transportation demand forecasts. This work aimed to perform a case study of aggregated transportation demand forecasts in 54 distribution centers of a Brazilian carrier. Methods: A computational simulation and case study methods were applied, exploring the characteristics of the datasets through autoregressive integrated moving average (ARIMA) and its variations, in addition to a deep neural network, long short-term memory, known as LSTM. Eight scenarios were explored while considering different data preprocessing methods and evaluating how outliers, training and testing dataset splits during cross-validation, and the relevant hyperparameters of each model can affect the demand forecast. Results: The long short-term memory networks were observed to outperform the statistical methods in ninety-four percent of the dispatching units over the evaluated scenarios, while the autoregressive integrated moving average modeled the remaining five percent. Conclusions: This work found that forecasting transportation demands can address practical issues in supply chains, specially resource planning management.

With the development of information technology and global economic integration, supply chain management has become a research hotspot in global management science. This article uses an electronic product as a supply chain management system. Based on the mathematical model and algorithm of supply chain design, the experiment designs an experiment based on traditional predictive models and optimized models to predict and analyze the sales of a certain electronic product. Experimental data shows that in order to adapt to the characteristics of supply chain demand forecasting in a big data environment, the proposed optimization model and method are reliable and practical. The experimental results show that the nine-month sales of the traditional prediction model are: 32, 38, 58, 57, 66, 73, 81, 90, 93, and the sales of the prediction model optimized by the cross genetic algorithm are: 41, 58, 72, 84, 95, 104, 103, 107, 92, to make up for the shortcomings of the model’s insufficient generalization ability under the high-dimensional situation, reduce the error of demand forecasting, and provide supply chain demand forecasting under the background of big data Effective forecasting methods have greatly alleviated the defects in supply chain management.KeywordsBig data backgroundLarge-scale supply chain designMathematical models and algorithms

Planning the future blood collection efforts must be based on adequate forecasts of transfusion demand. In this study, univariate time-series methods were investigated for their performance in forecasting the monthly demand for RBCs at one tertiary-care, university hospital. Three time-series methods were investigated: autoregressive integrated moving average (ARIMA), the Holt-Winters family of exponential smoothing models, and one neural-network-based method. The time series consisted of the monthly demand for RBCs from January 1988 to December 2002 and was divided into two segments: the older one was used to fit or train the models, and the younger to test for the accuracy of predictions. Performance was compared across forecasting methods by calculating goodness-of-fit statistics, the percentage of months in which forecast-based supply would have met the RBC demand (coverage rate), and the outdate rate. The RBC transfusion series was best fitted by a seasonal ARIMA(0,1,1)(0,1,1)(12) model. Over 1-year time horizons, forecasts generated by ARIMA or exponential smoothing laid within the +/- 10 percent interval of the real RBC demand in 79 percent of months (62% in the case of neural networks). The coverage rate for the three methods was 89, 91, and 86 percent, respectively. Over 2-year time horizons, exponential smoothing largely outperformed the other methods. Predictions by exponential smoothing laid within the +/- 10 percent interval of real values in 75 percent of the 24 forecasted months, and the coverage rate was 87 percent. Over 1-year time horizons, predictions of RBC demand generated by ARIMA or exponential smoothing are accurate enough to be of help in the planning of blood collection efforts. For longer time horizons, exponential smoothing outperforms the other forecasting methods.

Wind-sensitive structures usually suffer from violent vibrations or severe damages under the action of tropical storms. It is of great significance to forecast tropical-storm winds in advance for the sake of reducing or avoiding consequent losses. The model used for forecasting becomes a primary concern in engineering applications. This paper presents a performance evaluation of linear and nonlinear models for the short-term forecasting of tropical storms. Five extensively employed models are adopted to forecast wind speeds using measured samples from the tropical storm Rumbia, which facilitates a comparison of the predicting performances of different models. The analytical results indicate that the autoregressive integrated moving average (ARIMA) model outperforms the other models in the one-step ahead prediction and presents the least forecasting errors in both the mean and maximum wind speeds. However, the support vector regression (SVR) model has the worst performance on the selected dataset. When it comes to the multi-step ahead forecasting, the prediction error of each model increases as the number of steps expands. Although each model shows an insufficient ability to capture the variation of future wind speed, the ARIMA model still appears to have the least forecasting errors. Hence, the ARIMA model can offer effective short-term forecasting of tropical-storm winds in both one-step and multi-step scenarios.