Next Article in Journal
Energy Consumption Analysis Using Measured Data from a Net-Zero Energy Commercial Building in a Cold and Dry Climate
Previous Article in Journal
Fermentative Production of Erythritol from Cane Molasses Using Candida magnoliae: Media Optimization, Purification, and Characterization
Previous Article in Special Issue
Loss Aversion Order Strategy in Emergency Procurement during the COVID-19 Pandemic
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Intelligent Vehicle Sales Prediction Based on Online Public Opinion and Online Search Index

1
School of Economics and Management, Beijing Forestry University, Beijing 100083, China
2
School of Economics and Management, University of Science and Technology Beijing, Beijing 100083, China
*
Author to whom correspondence should be addressed.
Sustainability 2022, 14(16), 10344; https://doi.org/10.3390/su141610344
Submission received: 1 July 2022 / Revised: 15 August 2022 / Accepted: 17 August 2022 / Published: 19 August 2022

Abstract

:
Intelligent vehicles refer to a new generation of vehicles with automatic driving functions that is gradually becoming an intelligent mobile space and application terminal by carrying advanced sensors and other devices and using new technologies, such as artificial intelligence. Firstly, the traditional autoregressive intelligent vehicle sales prediction model based on historical sales is established. Secondly, the public opinion data and online search index data are selected to establish a sales prediction model based on online public opinion and online search index. Then, we consider the influence of KOL (Key Opinion Leader), a sales prediction model based on KOL online public opinion andonline search index is established. Finally, the model is further optimized by using the deep learning algorithm LSTM (Long Short-Term Memory network), and the LSTM sales prediction model based on KOL online public opinion and online search index is established. The results show that the consideration of the online public opinion and search index can improve the prediction accuracy of intelligent vehicle sales, and the public opinion of KOL plays a greater role in improving the prediction accuracy of sales than that of the general public. Deep learning algorithms can further improve the prediction accuracy of intelligent vehicle sales.

1. Introduction

Intelligent vehicles are a new generation of vehicles that are equipped with advanced sensors and other devices, use new technologies such as artificial intelligence, have self-driving functions, and gradually become intelligent mobile spaces and application terminals. In recent years, a new generation of industrial technology with intelligence and the Internet as the core has been on the rise, promoting traditional industries to accelerate the “intelligent” transformation and upgrade [1,2]. As one of the most widespread and important industries, the automotive industry is also accelerating its progress of intelligence, and “Intelligent vehicles” are receiving more attention, especially in the areas of shared mobility, energy consumption, and vehicle safety [3,4]. The majority of intelligent vehicles are electric vehicles, and the number of electric vehicles will continue to grow in the next 40 years with an “S” trend [5]. The Chinese government has also made intelligent vehicles a priority to the development of the automotive industry, and the development strategy released by the country has drawn strong attention to the industry. Intelligent vehicles are increasingly becoming the focus of the automotive industry, but research related to intelligent vehicles is still relatively lacking.
Predicting sales is a key step in making production decisions on companies and public policy for governments. Companies use product sales predictions as a basis forestimating sales revenues. They can also use product sales predictions to develop plans for marketing, sales management, production, purchasing, and logistics to improve economic efficiency and reduce losses in production planning. Intense competition, large investments, and the need for rapid model updates characterize the automotive industry, which makes predicting crucial for sale and production processes [6]. Most intelligent vehicle companies are emerging vehicle makers and have a low grasp of the market demand for their products, so it is more important to accurately predict the sales scale of intelligent vehicles for companies to set a reasonable production scale. In addition, as a strategic emerging industry supported by the Chinese government, the sales growth of intelligent vehicles is the result of the joint action between the market and the government. Therefore, the study of the intelligent vehicle market sales prediction is also important to the formulation of intelligent vehicle industry support policies and the arrangement of related supporting facilities.
Big data have become one of the important tools in the field of predicting [7] and it has been very widely used in the traditional automotive field. Unlike traditional vehicles, intelligent vehicle brands come with Internet properties and are often also referred to as Internet vehicles, which is why they are widely followed and discussed on the Internet; with more than 6.9 million discussions about Tesla (Fremont, CA, USA), the intelligent vehicle brand, on the social platform Twitter from 2018 to 2020. In China, NIO, the emerging domestic intelligent vehicle brand, has 910,000 followers on the social platform Weibo, with more than 200,000 discussions about NIO on Weibo alone from 2018 to 2020. In this case, the brand’s online public opinion and attention often affect the user’s willingness to purchase and further affect the sales of the vehicle brand. In 2019, the domestic intelligent vehicle brand XPeng had a big drop in sales due to a collective online rights opinion storm, with sales down 43.9%. Previous studies have tended to focus only on traditional vehicle brands and less on the impact of online public opinion and online search index on intelligent vehicle sales. With the growing development of intelligent vehicles, it is urgent to establish a sales prediction model that applies to intelligent vehicles and takes into account the influence of online public opinion and attention.
To address the above issues, this paper takes NIO and XPeng as examples, and based on the traditional sales prediction model based on historical sales, considers online public opinion sentiment and Baidu Index data, and especially analyzes the influence of KOL on online public opinion. The model is then further optimized by using the deep learning algorithm LSTM (Long Short-Term Memory network) to establish the LSTM sales prediction model based on KOL online public opinion and online search index. The LSTM sales prediction model based on KOL online public opinion and online search index to improve the accuracy of prediction for the sale of NIO by 60.37% and the MAPE (Mean Absolute Percentage Error) reached 9.7718% when compared to the traditional autoregressive model. The accuracy of prediction for the sale of XPeng is improved by 84.37% compared to the autoregressive model, and the MAPE reached 5.899%. In other words, this paper finds that considering online public opinion and the search index can greatly improve the prediction accuracy of intelligent vehicle brand sales, and the effect of KOL’s online public opinion on sale prediction accuracy is much greater than that of general public opinion, which also explains why ordinary consumers often feel that their voices on the Internet are not valued by enterprises. Through multiple case tests, the prediction model proposed in this paper can help enterprises optimize sales prediction accuracy and is important to them for setting a reasonable production scale.

2. Literature Review

2.1. Prediction Research Based on Brand Network Emotion

With the development of social media, people are happy to post their personal feelings on the Internet, and this online feedback is very useful to predict the sales volume trends of various products.
Herr, et al. [8] found that brands’ online public opinion sentiment (or Internet Word-of-Mouth) has a significant impact on consumers’ purchase behavior. Ren and Deng [9] found that online public opinion affects the RMB exchange rate. Yan, et al. [10] found that in addition to e-commerce platforms, social media electronic word-of-mouth reviews also influence consumers and that differences in platforms also lead to differences in consumer perceived usefulness. Research by Liu [11], Godes and Mayzlin [12], and Chevalier and Mayzlin [13] found that online public opinion (including the number of online reviews, ratings, and the content of reviews) influences consumer behavior and product sales. The empirical research of Huang, et al. [14] found that the stock market prediction model with the addition of Weibo sentiment was able to obtain better accuracy. Yu, et al. [15] trained a sentiment-based semantic analysis model to extract sentiment from online reviews and they found that sentiment information and the quality of online reviews had a significant impact on box office prediction. Archak, et al. [16] collected daily product prices, product ratings, and review opinions from Amazon consumer reviews and built a linear equation study to find that textual data from product reviews can be used to determine consumers’ relative preferences for different product features and thus predict future sales. Using multivariate linear autoregressive and neural network algorithms, Lv [17] proposed an autoregressive dimensional sentiment model and a positive and negative dimensional sentiment model to study the prediction of daily box office and first-week box office. The results confirmed that dimensional sentiment extracted from a multidimensional sentiment analysis framework can improve the accuracy and robustness of box office revenue prediction models. Jiang and Li [18] combined Weibo comment sentiment with economic indicators and historical efficiency to predict vehicle model sales, and the results showed that the inclusion of comment perception sentiment and macroeconomic indicators effectively improved sales prediction accuracy. Zhang, et al. [19] proposed a method for predicting car sales using online reviews and search engine data. By combining principal component analysis (PCA), the Back Propagation Neural Network (BPNN), and the improved fruit fly optimization algorithm (DSFOA), a prediction model PCA–DSFOA–BPNN is constructed, and the study finds that the model can effectively improve the prediction accuracy. Jiang, et al. [20] investigated the impact of investor sentiment on the forecasting ability of crude oil prices in China. The study shows that the long short-term memory model combined with the composite sentiment index performed the best, with a lower rate of prediction errors and greater accuracy. In summary, it can be seen that previous research on the prediction of online sentiment has mostly focused on everyday consumer products, with less attention paid to emerging intelligent vehicle brands.

2.2. Prediction Research Based on Online Search Index

The current authoritative online search index is Google Trends and Baidu Index; the difference between the two mainly lies in the fact that Baidu has a larger Chinese user base than Google and can provide more accurate data references to Chinese issues [21]. Geva, et al. [6] used Google Trends data to improve the effectiveness of predicting vehicle sales. Choi and Varian [22] used Google Trends data to predict the sales of different vehicle brands in the U.S. The baseline AR model (Autoregressive model) was compared to two models that included Google Trends variables to give prediction results for three brands: Chevrolet, Toyota, and Ford. The study found that adding Google Trends data can improve the measured value by about 10.5%. Boone, et al. [23] found that introducing Google Trends data into the sales prediction model can reduce out-of-sample prediction errors. Huang, et al. [24] analyzed the relationship between Internet search data and actual tourist traffic using the covariance theory and Granger causality. The study showed that there is a long-run equilibrium relationship and Granger causality between the number of tourists and a set of related keywords in the Baidu Index, which indicates that an increase in the Baidu keyword search index is positively related to the observed increase in tourist traffic. Li, et al. [25] established a CLSI-EMD-BP model based on Internet searches, and the experimental results surface that the prediction error of this model is significantly lower than the three benchmark models of time series, Internet searches, and BP neural networks. Liu, et al. [26] concluded that the search index has significant predictive power for the annual return of the SSE index by the Granger causality test. Kang, et al. [27] explored the mapping relationship between public Internet search behavior and real travel behavior through the Granger causality test and ARIMA model, and the results showed that the geographic location attribute of the Baidu Index helps to improve the prediction accuracy. Furthermore, the empirical research of Wang, et al. [28] concluded that there is a significant positive effect of Internet searches on vehicle sales. Jiang, et al. [29] proposed a consumer attention metric integrating word-of-mouth reviews and search data to predict vehicle sales, and the results showed that the model with the introduction of consumer attention reduced the RMSE (Root Mean Squared Error) and MAPE metrics by 2.02 and 0.96%, respectively. Yao, et al. [30] proposed a hybrid forecast model of daily tourist arrivals based on the Baidu Index, which couples rescaled range analysis (R/S), support vector regression (SVR), and autoregressive integrated moving average (ARIMA). The study showed that the Baidu Index helps to improve the accuracy of predicting visitor numbers. Jin, et al. [31] proposed a novel hybrid model embedded with the Baidu Search Index to achieve multi-step metro passenger flow prediction. The empirical results show that the proposed model can significantly outperform the benchmark model in terms of both the level and directional accuracy.

2.3. Prediction Research Based on Deep Learning

Machine learning techniques and deep learning algorithms introduce new approaches to the prediction problem so that the relationships between variables are modeled in a deep hierarchy. In recent years, algorithm long- and short-term memory (LSTM) based on deep learning has been widely used. LSTM has been used for time series predicting [32,33,34], as well as for economic and financial data, such as predicting the volatility of the S&P 500 [35]. LSTM is a special case of the recurrent neural network (RNN) approach, originally proposed by Hochreiter and Schmidhuber [33], which is a relatively new approach to solving prediction problems. Through an empirical study, Ouyang, et al. [36] found that they were able to effectively predict the long- and short-term dynamic trends of financial time series data. Chang [37] used the ARIMA model to fit the sales data and used it as the initial connection value of the RNN to build the ARIMA–RNN model. Liu, et al. [38] designed a convolutional neural network-based vehicle sales prediction model, which has higher prediction accuracy compared to the RBF model, ARIMA model, and ARIMA–RBF hybrid model. Liu, et al. [39] proposed an online public opinion prediction method integrating Weibo hotspot analysis and the LSTM model and confirmed that the model has a good prediction effect. Kong and Zhu [40] investigated the approach of adaptive input selection (AIS) for the trend prediction of high-frequency stock index prices, using two state-of-the-art machine learners, a support vector machine (SVM), and an artificial neural network (ANN) as learning models. It is shown that the AIS approach using t-statistics, information gain, and ROC methods can achieve better prediction performance than the deterministic input setting (DIS) approach. Pamuksuz, et al. [41] developed a hybrid machine learning algorithmic design (LDA2Vec) to predict competitors’ brand personalities, and the approach provides practitioners with the ability to foster branding strategies by using big data resources. Alantari, et al. [42] found that neural network-based machine learning methods, in particular pre-trained versions, offer the most accurate predictions, while topic models such as Latent Dirichlet Allocation offer deeper diagnostics. Fewer scholars have applied deep learning techniquestion telligent vehicle sales predicting.

2.4. Research on Vehicle Sales Predicting

Vehicle sales predicting are divided into two main categories, one using traditional data for predicting and the other adding big data to traditional analysis for predicting. Traditional data mainly include macro factors and internal data of vehicle companies. Zhao [43] studied the main factors affecting vehicle sales, including economy, price, and environment, and considered government policies, transportation infrastructure, and other influencing variables. Chen [44] used historical auto sales data and an autoregressive integrated moving average model (ARIMA) to predict the demand for the entire Chinese market, and the experimental results showed that the ARIMA model was effective in predicting macro auto sales. In recent years, more and more studies have taken into account the impact on big data. Cui [45] revealed the relationship between Internet search data and vehicle sales, and the analysis and testing results showed that the method has a high prediction accuracy compared to traditional methods of predicting vehicle sales. Liu, et al. [46] discussed the effect of historical sales and brand sentiment on vehicle sales predicting using online review data and sales data, and the results showed that the average prediction error of the prediction model was 5.93%, which was 6.24% lower than that of the ARIMA model. Zhang [47] introduced IWOM and search index in the prediction process of traditional vehicles, and the results showed that the introduction of search data based on the IWOM model could improve the prediction accuracy and effectively improve the prediction of vehicle sales. Wang, et al. [48] utilized a panel vector autoregressive model to investigate the offline car sales prediction. The result indicates that incorporating firm-generated content and user-generated content from Facebook, Twitter, and traditional media could improve the performance of offline sales prediction. Liu, et al. [49] proposed a multi-factor prediction model integrating DWT (Discrete Wavelet Transform) and BiLSTM (Bidirectional Long Short-Term Memory) to forecast New Energy Vehicle sales in China from 2020 to 2025; the study showed that the New Energy Vehicle market penetration rate could not reach the target of 20% in 2025.
In summary, we have compiled previous studies, as shown in Table 1. The use of online public opinion and online search index has been widely studied in the field of traditional vehicle sales predicting, but there is a relative lack of research on intelligent vehicle sales predicting, especially for the application of Internet factors, such as online public opinion and online search index. This paper will address this aspect and further enrich the related research.

3. Model

Using historical sales data to predict future demand is a common predicting tool, and Chen [44] used historical auto sales data to predict demand for the Chinese market. The flowchart of the study in this paper is shown in Figure 1.
The meanings of the symbolic representations of the variables used for this section are shown in Table 2.
Most of the vehicle sales predictions nowadays are based on historical sales data, exploring the annualized trend of vehicle sales and then making future sales predictions. Figure 2 shows the change in vehicle sales for NIO and XPeng.
The q-order autoregressive sales prediction model is first developed as the baseline model, where y t denotes the sales volume of month t and ε t is the error term. Due to the unpredictability of unexpected events and policy documents issued at different time points, e.g., the outbreak of a new coronavirus in 2020 and the frequent release of favorable policies of intelligent vehicles and new energy vehicles, resulting in a steep decrease or steep increase in sales data, a time dummy variable, d t , is introduced into the model in this paper to eliminate the influence of these factors.
y t = i = 1 q α i y t i + d t + ε t
Herr, et al. [8] found that a brand’s online public opinion has an important influence on consumers’ purchasing behaviors. Intelligent vehicles are born based on the Internet and come with Internet genes, so they are more likely to receive the influence of Internet factors. Therefore, this paper applies big data technology based on the sales prediction model based on historical sales to obtain a more accurate intelligent vehicle sales prediction model. The mathematical formula is expressed by adding the influence of Weibo sentiment ( j = 1 p β j w t j ) and Baidu Index ( k = 1 m γ k b t k ) on the sales volume based on Equation (1) to obtain a sales prediction model that considers both online public opinion and online search index, where w t denotes the online opinion index in period t and b t represents the Baidu Index in period t.
y t = i = 1 q α i y t i + j = 1 p β j w t j + k = 1 m γ k b t k + d t + ε t
In 2020, Li Auto cooperated with many KOLs short video bloggers who promoted Li Auto as the owner of “Li-One”, which created a good reputation and popularity for Li Auto and also drove the sales of Li Auto. According to Li, et al. [50], influential people can influence product sales. KOL include those who express their views on a certain field and have a certain influence. It is clear from past studies and actual marketing strategies of companies that KOLs have a significant impact on the sales of intelligent vehicles. However, in the previous model, the statements of KOLs and the general public were given the same weight and added together, which could not reflect the influence of KOLs and could not verify whether KOL’s statements were more relevant for sales prediction in the study.
Therefore, this paper further improves based on Equation (2) by removing the sentiment of the general public and considering only the sentiment of KOL’s Weibo, i.e., considering only Weibo with more than 50 likes or comments, which will be recorded as getting the sales prediction model considering KOL’s online public opinion and online search index.
y t = i = 1 q α i y t i + j = 1 p β j w k t j + k = 1 m γ k b t k + d t + ε t
Based on the traditional sales prediction model with the introduction of KOL’s online public opinion and online search index, the prediction accuracy of the improved model may still be low because the base model uses a linear model, so we will use the deep learning algorithm LSTM to optimize the model at a deeper level to obtain an intelligent vehicle sales prediction model with a high prediction accuracy.

4. Data and Experiment

4.1. Data Acquisition and Pre-Processing

In this paper, we will use the datasets of intelligent vehicle brands, NIO and XPeng, for experimental operations.

4.1.1. Public Opinion Data of Weibo

Since online communities began to flourish, many social platforms for gathering online opinions have emerged from the domestic Internet, such as Sina Weibo, Douban, and Zhihu. Sina Weibo is a platform based on building user relationships to share, disseminate, and receive information. Through the website or app, users can publicly upload images and videos for real-time sharing, while other users can post comments using text, images, and videos, or use the private messaging service.
In this paper, we use the public opinion dataset of “NIO” and “XPeng” from 1January 2018 to 31 March 2021 on Sina Weibo (hereinafter referred to as “Weibo”) as the source of online public opinion data. The total number of opinions of NIO is 136,010, including 5182 KOL opinions, accounting for 3.81%, while the total number of opinions of XPeng is 80,147, including 2550 KOL opinions, accounting for 3.18%. Weibo is chosen as the source of online public opinion text data for the following reasons:
  • Weibo has a wide range of users and is highly active. Weibo has more than 500 million registered users; among them, there are 313 million active users every month, and more than 100 million messages posted by users every day, which makes the public opinion content-rich and mostly reliable, and the public opinion data are representative.
  • The sales volume and Baidu Index data used in this paper are all domestic Chinese data. Compared to global social platforms such as Twitter, the main user group of Weibo is Chinese users, its content can reflect the emotional tendency towards Chinese netizens, and the results are more accurate.
Sentiment analysis refers to the use of natural language processing, text analysis, and linguistics to form valuable insights from data to help people make decisions [51]. Sentiment analysis allows for the bulk extraction of sentiment tendencies expressed by authors of Weibo and converts unstructured Weibo texts into structured sentiment indices for introduction into the model. In this paper, we use Harbin Institute of Technology’s pyLTP (pyLTP is a Python wrapper for the Language Technology Platform developed by Harbin Institute of Technology based on the CNKI Sentiment Dictionary. It provides functions such as word division, lexical annotation, named entity recognition, dependent syntactic analysis, and semantic role annotation, a natural language processing toolkit for sentiment analysis of Weibo.
In this paper, we use the pyLTP tool to split each Weibo text into sentences and compare the results with the discontinued words list of Harbin Institute of Technology to remove the discontinued words.
Sentiment analysis is performed on each sentence after the clause. Use the CNKI Sentiment Dictionary to determine the number of sentiment words within a sentence. If there are positive words, the positive sentiment index increases by 1, and if there are negative words, the negative sentiment index increases by 1. Adverbs of degree will affect the emotional degree of the above emotional words, so it cannot simply add 1 or subtract 1. First of all, it is necessary to judge whether there is an adverb of degree. If there is, multiply the corresponding emotional word by the corresponding degree coefficient and then, add it to the emotional value; the existence of negative words will make the emotional words express the opposite emotion, so a negative coefficient should be given. Calculate the sentiment value of each clause and then, sum the sentiment values of all clauses to obtain the sentiment value of each Weibo post.
When using the above approach from sentiment analysis, some Weibo posts with long content tends to obtain large positive or negative sentiment values, and this extreme Weibo has a great influence weight if monthly aggregation is performed directly. In this paper, each Weibo is equally weighted in the monthly sentiment values by default. To ensure that each Weibo is equally weighted, this paper ignores the specific sentiment value of each Weibo and only retains its sentiment tendency. The sentiment value of each Weibo post was recorded as 1 (positive sentiment), −1 (negative sentiment), and 0 (neutral sentiment) according to its final sentiment tendency. The monthly emotional value of Weibo was summed up to obtain the monthly emotional value.
After the sentiment analysis based on the CNKI Sentiment Dictionary, the results were obtained: 91,957 Weibo texts had a positive sentiment tendency (recorded as 1), 38,113 Weibo texts had a negative sentiment tendency (recorded as −1), and 5940 texts had a neutral sentiment tendency (recorded as 0). The monthly sentiment values were obtained by summing up the monthly Weibo sentiment, some of which are shown in Table 3. Detailed data are given in the table.

4.1.2. Sales Volume Data

Vehicle sales data were obtained from the Home of Owners website. NIO and XPeng were established and sold for a short period of time. Most automotive websites do not have complete sales data for NIO; however, the Home of Owners website retains complete monthly sales data for NIO from July 2018 to date, and XPeng from December 2018 to date. Table 4 shows some of the NIO sales data from the Home of Owners website from July 2018 to March 2021. Detailed data are given in the table.

4.1.3. Baidu Index

The sales volume studied in this paper is the sales volume in mainland China, so the Baidu Index is chosen as the source of Internet search index data. The Baidu Index data for the keywords “NIO” and “XPeng” were obtained from 1 January 2018 to 31March 2021, and some of the data are shown in Table 5.
The obtained Baidu Indexes are daily data, and since the subsequent modeling predictions will be modeled monthly, the daily Baidu Indexes need to be transformed into monthly Baidu Indexes. The daily Baidu Index data within a month are summed to produce the partial monthly Baidu Index data shown in Table 6. Detailed data are given in the table.

4.2. Experiment Settings

This paper uses Eviews software (Eviews10.0, IHS Global INC., Irvine, CA, USA) to model Equations (1)–(3). MAPE is used as the evaluation index of the model.
M A P E = 100 % n i = 1 n | y ^ i y i y i |
where y ^ i represents the estimated value and y i represents the actual value. A smaller MAPE value indicates better model prediction accuracy.
A deep learning model is built using the Pytorch (Pytorch1.8, Facebook AI Research, CA, USA) framework, and MAPE is used as the evaluation metric for the model. The training process is as follows:
  • Pre-processing of data. The results fitted by the sales prediction model based on KOL online public opinion and online search index cannot be directly used as the input of LSTM and need to be normalized. Therefore, use 70% of the data as the training set and 30% as the test set.
  • Initialization of parameters. Rational configuration of the hyper-parameters of LSTM.
  • Determine the gradient descent algorithm. The model is trained using Adam’s algorithm.
  • Training and output. Use the arithmetic power of the Google Colab platform to train the model and make predictions, output the predictions and perform the inverse normalization operation, output the predicted values, and calculate MAPE.
In this paper, the Pytorch deep learning framework is used to train the model, and the model parameters are set as shown in Table 7.

4.3. Experimental Results and Analysis

As can be seen, more satisfactory prediction accuracy is obtained using the LSTM algorithm and it is further reduced by 40.3% when compared to the sales prediction model based on KOL online public opinion and online search index. Compared with the traditional sales prediction model based on historical sales, it is 60.37% lower and has a better prediction effect.
From Table 8 and Figure 3 and Figure 4, it can be seen that the sales prediction model based on historical sales can predict the long-term sales trend, and due to the introduction of time dummy variables, it can also predict well for sudden changes after 2020. The model’s prediction results are relatively smooth and it cannot make accurate predictions for short-term sales fluctuations, which cannot provide a reference for short-term strategy adjustment of enterprises. The main reasons are that the important source of short-term fluctuations in sales volume is the influence of news and events, and the model does not consider related factors, so it cannot predict short-term fluctuations. This in turn leads to too high MAPE values and poor overall predicting accuracy.
The sales prediction model based on online public opinion and online search index is not only able to predict the long-term sales predict trend but also, due to the introduction of factors such as online public opinion and online search index, the influence of short-term events and news is incorporated into the model, which has a good prediction effect on short-term sales fluctuations, especially for the prediction of sales fluctuations after 2020, with a MAPE value of 20.35157, which is 4.30851 lower compared to the sales prediction model based on historical sales. However, the model is not very effective against predicting short-term sales fluctuations before 2020. Considering that the intelligent vehicle field has just started before 2020, the relevant online public opinions are mainly published by ordinary consumers, and the comments are relatively common and scattered, which have little impact on vehicle sales.
The sales prediction model based on KOL online public opinion and online search index is not only able to predict the long-term sales prediction trend and sales fluctuation after 2020, but also, due to the emphasis on the role of KOL, the short-term sales fluctuation before 2020 can be predicted well. The overall trend and fluctuation of the predicted value are generally consistent with the real data, and the MAPE value is reduced to 16.36934. The MAPE value is also reduced to 16.36934, which is 8.29074 lower than the sales prediction model based on historical sales, and 3.98223 lower than the sales prediction model based on online public opinion and online search index.
After combining the previous sales prediction model based on KOL online public opinion and online search index with the deep learning algorithm, the model can not only fit the long-term sales trend and short-term fluctuations as well as the previous model, but can also further optimize the prediction effect and receive a relatively good prediction effect. In addition, we observed that the autoregressive model performs better on the NIO dataset, while the LSTM-based model performs better on the XPeng dataset. The reason for this difference is that the sales data of XPeng are more volatile, and the timeseries regression method is relatively smoother, while the deep learning method learns this information more easily. Therefore, after using the deep learning method, the XPeng with more fluctuating data received better prediction results.

5. Discussion and Suggestions

The traditional sales prediction model based on historical sales can only predict the long-term trend, but the curve is relatively smooth, the prediction effect for short-term sales fluctuations is average, and the short-term reference significance is not great. The introduction of online public opinion and online search index can improve the prediction accuracy and has a good effect on the prediction of short-term fluctuations in the later period, indicating that online public opinion and online search index can improve the accuracy of sales prediction by reflecting events and news trends on the network. However, the prediction effect of short-term fluctuations in the early stage is still not obvious. In the part of online public opinion, the model considering only KOL can further improve the prediction accuracy and has a good effect on the prediction accuracy of short-term fluctuations, which indicates that the microblogs of KOL are more able to reflect the big trends of events and news and have a better effect on improving the prediction.
Considering online public opinion and online search index can greatly improve the prediction accuracy of intelligent vehicle brand sales. After considering online public opinion and online search index, the prediction accuracy of the model improves by 17.5% compared with the prediction model based on historical sales. In addition, among the online public opinions, KOL’s public opinion has a much greater effect on the accuracy of sales prediction than the general public’s public opinion. Considering only KOL’s online public opinion, the prediction accuracy of the model has further improved by 33.6% compared to the prediction model based on historical sales. Thus, it can be seen that, among the influence of online public opinion, the general public opinion plays a small role in improving sales prediction, and it is mainly the influence of KOL. This is the fundamental reason why ordinary consumers often feel that their voices on the Internet are not valued by companies.
Based on the above conclusions, the subsequent research and application of sales prediction of intelligent vehicles should focus on the influence of online public opinion and online search index, especially the influence of online public opinion published by KOL on sale.
For companies, we make the following recommendations:
  • Pay attention to the brand network image.
From the above experimental results, we can see that online public opinion and online search index can greatly influence intelligent vehicle sales, and intelligent vehicle manufacturers should make efforts to build a good network image on the Internet. On the one hand, it is important to maintain a direct network image. Companies should actively market through the network to increase the network exposure of their highlights and selling points as much as possible to gain more consumers’ attention. On the other hand, we should maintain all the factors that indirectly affect the network image and provide consumers with good after-sales service and customer service, so that consumers can receive a pleasant experience to indirectly improve the brand’s network image.
  • Use the influence of KOL for marketing.
From the above experimental results, we can see that the influence of KOL’s online public opinion on sale is much greater than that of the general public. On the one hand, enterprises can grasp the large and let go of the small and can focus their resources on KOL’s marketing in the process of conducting online marketing. For example, the cooperation between Li Auto and many Jitterbug netizens is one of the successful cases of applying KOL marketing in the field of intelligent vehicles. Using KOL influences the majority of consumers to improve the overall network image, and thus promotes sales. On the other hand, companies can cultivate and replicate more KOLs themselves to obtain more stable marketing effects, such as Starbucks’ internal incubation of “Lee No Latte” and L’Oréal’s internal incubation of “Li Jiaqi”. Internally incubated KOLs have a higher brand loyalty and deeper cooperation, can produce better content for the company, and improve the company’s online image.
  • Pay attention to the sales predictions and rationalize the production plan.
From the experimental results above, we can see that NIO’s sales have had great fluctuations from 2018 to 2020. If we do not make sales predictions and blindly produce, it will bring huge inventory and capital chain problems. Therefore, after building a series of models, we can predict long-term trends and short-term fluctuations more accurately and provide a reference for the enterprise’s production plan development. Most intelligent vehicle enterprises are startups with relatively inexperienced and fragile capital chains. They should actively use information systems to predict sales volume and make reasonable production plans based on the predicted results. In the research and application of sales predicting of intelligent vehicles, various deep learning algorithms should be actively considered. For enterprises, they should also choose deep learning algorithms that fit their situation according to their reality and train their models based on their data to make more accurate sales predictions and provide references for further rational production arrangements.

6. Conclusions

Based on an in-depth study of prediction theory, this paper proposes an LSTM sales prediction model based on KOL online public opinion and online search index by combining the Internet attributes of intelligent vehicles and shows a good prediction effect on the sales prediction of intelligent vehicles. It is also validated by replacing the dataset, which proves the good generalization of the model.
The model proposed in this paper can help companies optimize the accuracy of sales forecasts and is a guide to setting a reasonable production scale. The limitations of this study are mainly reflected in the following aspects: (1) This paper only considers the effect of Weibo content text on sales prediction. As an open and interactive social platform, Weibo’s comments, likes, and reposts data can further reflect the public’s sentiment towards specific content. In the next step, Weibo comments, likes, and reposts will be taken into account in the prediction model. (2) This paper only considers the sentiment of the Weibo textual content. The emoji expressions and pictures deleted in the pre-processing stage are more likely to express the publisher’s sentiment, and the next step will be to consider including the above content in the scope of sentiment analysis. (3) The factors that affect intelligent vehicle sales are complex, such as policies, gas and electricity prices, and loan rates. In this paper, only the effects of online public opinion and online search index are considered, and the next step is to fully consider various factors and further improve the prediction model. (4) The ensemble learning method can improve the accuracy of model prediction and will be applied in future studies.

Author Contributions

Conceptualization, M.Z.; methodology, M.Z. and X.P.; software, H.X.; validation, M.Z. and H.X.; formal analysis, M.Z. and X.P.; investigation, N.M.; resources, M.Z. and H.X.; data curation, H.X. and X.P.; writing—original draft preparation, M.Z. and H.X.; writing—review and editing, M.Z., H.X. and N.M.; visualization, H.X. and X.P.; supervision, M.Z. and N.M.; project administration, M.Z.; funding acquisition, M.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the National Natural Science Foundation of China (Grant No. 71901027), Beijing Forestry University 2021 Curriculum Ideological and Political Project (Grant No. 2021KCSZXY010) and the Supply Chain Management Teaching Program (Grant No. KCSZ22008).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The corresponding author can be contacted if there is the need to further cross-examine the data used.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Miao, X. Smart Factory and the Transformation & Upgrading of Equipment Manufacturing Industry. Process Autom. Instrum. 2014, 35, 1–6. [Google Scholar]
  2. Pei, C.; Yu, Y. New development of German “industry 4.0” and Sino German manufacturing cooperation. Res. Financ. Econ. Iss. 2014, 10, 27–33. [Google Scholar]
  3. TuroŃ, K.; Czech, P.; TÓTh, J. Safety and Security Aspects in Shared Mobility Systems. Sci. J. Sil. Univ. Technol. 2019, 104, 169–175. [Google Scholar] [CrossRef]
  4. Turoń, K.; Kubik, A.; Chen, F. Operational Aspects of Electric Vehicles from Car-Sharing Systems. Energies 2019, 12, 4614. [Google Scholar] [CrossRef]
  5. Guo, D.; Zhang, H.; Zheng, C.; Gao, S.; Wang, D. Analysis of the future development of Chinese auto energy saving and environmental benefits. Syst. Eng. Theor. Pract. 2016, 36, 1593–1599. [Google Scholar]
  6. Geva, T.; Oestreicher-Singer, G.; Efron, N.; Shimshoni, Y. Using Forum and Search Data for Sales Prediction of High-Involvement Products. MIS Quart. 2017, 41, 65–82. [Google Scholar] [CrossRef]
  7. Chauhan, S.; Agarwal, N.; Kar, A.K. Addressing big data challenges in smart cities: A systematic literature review. Info 2016, 18, 73–90. [Google Scholar] [CrossRef]
  8. Herr, P.M.; Kardes, F.R.; John, K. Effects of Word-of-Mouth and Product-Attribute Information on Persuasion: An Accessibility-Diagnosticity Perspective. J. Consum. Res. 1991, 17, 454–462. [Google Scholar] [CrossRef]
  9. Ren, X.; Deng, L. Effect of Network Public Opinion Shocks on Exchange Rate: Evidence from the China-US Trade Friction. J. Manag. Sci. 2019, 32, 46–56. [Google Scholar]
  10. Yan, Q.; Ma, L.; Wu, S. The Influence of the eWOM Platforms’ Differences on Consumers’ Perceived Usefulness. J. Manag. Sci. 2019, 32, 80–91. [Google Scholar]
  11. Liu, Y. Word of Mouth for Movies: Its Dynamics and Impact on Box Office Revenue. J. Mark. 2006, 70, 74–89. [Google Scholar] [CrossRef]
  12. Godes, D.; Mayzlin, D. Using Online Conversations to Study Word-of-Mouth Communication. Mark. Sci. 2004, 23, 545–560. [Google Scholar] [CrossRef]
  13. Chevalier, J.A.; Mayzlin, D. The Effect of Word of Mouth on Sales: Online Book Reviews. J. Mark. Res. 2006, 43, 345–354. [Google Scholar] [CrossRef]
  14. Huang, R.; Zuo, W.; Bi, L. Predicting the Stock Market Based on Microblog Mood. J. Ind. Eng. Eng. Manag. 2015, 29, 47–52. [Google Scholar]
  15. Yu, X.; Liu, Y.; Huang, X.; An, A. Mining Online Reviews for Predicting Sales Performance: A Case Study in the Movie Domain. IEEE Trans. Knowl. Data Eng. 2012, 24, 720–734. [Google Scholar] [CrossRef]
  16. Archak, N.; Ghose, A.; Ipeirotis, P.G. Deriving the Pricing Power of Product Features by Mining Consumer Reviews. Manag. Sci. 2011, 57, 1485–1509. [Google Scholar] [CrossRef]
  17. Lv, X. Movie Box Office Prediction based on Multidimensional Sentiment Analysis in Social Media Environment. Ph.D. Thesis, Hefei University of Technology, Hefei, China, 2019. [Google Scholar]
  18. Jiang, K.; Li, Y. Forecast of Automobile Sales Considering the Perceived Emotion of Online Reviews and the Macroeconomic Indicators. Sci. Technol. Eng. 2021, 21, 5847–5854. [Google Scholar]
  19. Zhang, C.; Tian, Y.; Fan, Z. Forecasting sales using online review and search engine data: A method based on PCA–DSFOA–BPNN. Int. J. Forecast. 2022, 38, 1005–1024. [Google Scholar] [CrossRef]
  20. Jiang, Z.; Zhang, L.; Zhang, L.; Wen, B. Investor sentiment and machine learning: Predicting the price of China’s crude oil futures market. Energy 2022, 247, 123471. [Google Scholar] [CrossRef]
  21. Vaughan, L.; Chen, Y. Data mining from web search queries: A comparison of google trends and baidu index. J. Assoc. Inf. Sci. Technol. 2015, 66, 13–22. [Google Scholar] [CrossRef]
  22. Choi, H.; Varian, H.A.L. Predicting the Present with Google Trends. Econ. Rec. 2012, 88, 2–9. [Google Scholar] [CrossRef]
  23. Boone, T.; Ganeshan, R.; Hicks, R.L.; Sanders, N.R. Can Google Trends Improve Your Sales Forecast? Prod. Oper. Manag. 2018, 27, 1770–1774. [Google Scholar] [CrossRef]
  24. Huang, X.; Zhang, L.; Ding, Y. The Baidu Index: Uses in predicting tourism flows–A case study of the Forbidden City. Tour. Manag. 2017, 58, 301–306. [Google Scholar] [CrossRef]
  25. Li, X.; Lv, B.; Zeng, P.; Liu, J. Tourism prediction using web search data based on CLSI-EMD-BP. Syst. Eng.-Theory Pract. 2017, 37, 106–118. [Google Scholar]
  26. Liu, Y.; Lv, B.; Peng, G. Predictive Power of Internet Search Data for Stock Market: A Theoretical Analysis and Empirical Test. Bus. Manag. J. 2011, 33, 172–180. [Google Scholar]
  27. Kang, J.; Guo, X.; Fang, L. Tourism Trend Prediction Based on Baidu Index Spatial and Temporal Distribution. J. Southwest China Norm. Univ. (Nat. Sci. Edit.) 2020, 45, 72–81. [Google Scholar]
  28. Wang, L.; Ning, Y.; Jia, J. Predicting Sales and Market Share with Online Search: Evidence from Chinese Automobile Market. J. Ind. Eng. Eng. Manag. 2015, 29, 56–64. [Google Scholar]
  29. Jiang, C.; Wang, X.; Wang, Z. Forecasting Car Sales Based on Consumer Attention. Data Anal. Knowl. Discov. 2021, 5, 128–139. [Google Scholar]
  30. Yao, L.; Ma, R.; Wang, H. Baidu index-based forecast of daily tourist arrivals through rescaled range analysis, support vector regression, and autoregressive integrated moving average. Alex. Eng. J. 2021, 60, 365–372. [Google Scholar] [CrossRef]
  31. Jin, K.; Sun, S.; Li, H.; Zhang, F. A novel multi-modal analysis model with Baidu Search Index for subway passenger flow forecasting. Eng. Appl. Artif. Intell. 2022, 107, 104518. [Google Scholar] [CrossRef]
  32. Gers, F.A.; Schmidhuber, J.; Cummins, F. Learning to Forget: Continual Prediction with LSTM. Neural Comput. 2000, 12, 2451–2471. [Google Scholar] [CrossRef] [PubMed]
  33. Hochreiter, S.; Schmidhuber, J. Long Short-Term Memory. Neural Comput. 1997, 9, 1735–1780. [Google Scholar] [CrossRef] [PubMed]
  34. Schmidhuber, J. Deep learning in neural networks: An overview. Neural Netw. 2015, 61, 85–117. [Google Scholar] [PubMed]
  35. Huck, N. Pairs selection and outranking: An application to the S&P 100 index. Eur. J. Oper. Res. 2009, 196, 819–825. [Google Scholar]
  36. Ouyang, H.; Huang, K.; Yan, H. Prediction of Financial Time Series Based on LSTM Neural Network. Chin. J. Manag. Sci. 2020, 28, 27–35. [Google Scholar]
  37. Chang, D. Research on Automobile Sales Forecasting Model of ARIMA-RNN Based on Sentiment Analysis. Master’s Thesis, Lanzhou University of Finance and Economics, Lanzhou, China, 2020. [Google Scholar]
  38. Liu, J.; Zhang, M.; Peng, H.; Jia, X. Automobile Sales Forecasting Model Based on Convolutional Neural Network. Comput. Sci. 2021, 48, 178–183. [Google Scholar]
  39. Liu, D.; Shen, Y.; Zhan, T.; Liu, Y.; Ying, Y. Network public opinion forecasting method fusing microblog hotspot analysis and LSTM model. J. Jiangsu Univ. (Nat. Sci. Edit.) 2021, 42, 546–553. [Google Scholar]
  40. Kong, A.; Zhu, H. Predicting Trend of High Frequency CSI 300 Index Using Adaptive Input Selection and Machine Learning Techniques. J. Syst. Sci. Inform. 2018, 6, 120–133. [Google Scholar] [CrossRef]
  41. Pamuksuz, U.; Yun, J.T.; Humphreys, A. A Brand-New Look at You: Predicting Brand Personality in Social Media Networks with Machine Learning. J. Interact. Mark. 2021, 56, 55–69. [Google Scholar] [CrossRef]
  42. Alantari, H.J.; Currim, I.S.; Deng, Y.; Singh, S. An empirical comparison of machine learning methods for text-based sentiment analysis of online consumer reviews. Int. J. Res. Mark. 2022, 39, 1–19. [Google Scholar] [CrossRef]
  43. Zhao, Y. The Reaserch of Car Sales Forecast Models in China Based on Regression Analysis. Master’s Thesis, Central China Normal University, Wuhan, China, 2014. [Google Scholar]
  44. Chen, D. Chinese automobile demand prediction based on ARIMA model. In Proceedings of the International Conference on Biomedical Engineering & Informatics, Shanghai, China, 15–17 October2011. [Google Scholar]
  45. Cui, D. An Empirical Study of Automobile Sale Forecast under the Background of Big Data. Master’s Thesis, Henan University, Kaifeng, China, 2014. [Google Scholar]
  46. Liu, Y.; Zhang, X.; Wang, J. An Automobile sales Prediction model incorporating brand opinion. J. Hefei Univ. Technol. (Nat. Sci.) 2017, 40, 1276–1282. [Google Scholar]
  47. Zhang, J. Research on Automobile Sales Forecast Based on Internet Word-of-Mouth and Search Data. Master’s Thesis, Qingdao University, Qingdao, China, 2020. [Google Scholar]
  48. Wang, Y.Y.; Guo, C.; Susarla, A.; Sambamurthy, V. Online to Offline: The Impact of Social Media on Offline Sales in the Automobile Industry. Inf. Syst. Res. 2021, 32, 582–604. [Google Scholar] [CrossRef]
  49. Liu, B.; Song, C.; Wang, Q.; Zhang, X.; Chen, J. Research on regional differences of China’s new energy vehicles promotion policies: A perspective of sales volume forecasting. Energy 2022, 248, 123541. [Google Scholar] [CrossRef]
  50. Li, X.; Xu, Y.; Sheng, P. Key people and key opinion leaders. Enterp. Manag. 2005, 2, 32–35. [Google Scholar]
  51. Shukri, S.E.; Yaghi, R.I.; Aljarah, I.; Alsawalqah, H. Twitter Sentiment Analysis: A Case Study in the Automotive Industry. In Proceedings of the Applied Electrical Engineering & Computing Technologies, Amman, Jordan, 3–5 November 2015. [Google Scholar]
Figure 1. Study flowchart.
Figure 1. Study flowchart.
Sustainability 14 10344 g001
Figure 2. Sales volume of NIO and XPeng.
Figure 2. Sales volume of NIO and XPeng.
Sustainability 14 10344 g002
Figure 3. Comparison of the prediction effect of the four models in NIO dataset.
Figure 3. Comparison of the prediction effect of the four models in NIO dataset.
Sustainability 14 10344 g003
Figure 4. Comparison of the prediction effect of the four models in XPeng dataset.
Figure 4. Comparison of the prediction effect of the four models in XPeng dataset.
Sustainability 14 10344 g004
Table 1. Literature review.
Table 1. Literature review.
ArticleOnline Public OpinionOnline Search IndexKOL Online Public OpinionVehicleIntelligent VehicleAutoregressive ModelLSTMDeep Learning
Herr, et al. (1991) [8]
Ren and Deng (2019) [9]
Liu (2006) [11]
Godes and Mayzlin (2004) [12]
Chevalier and Mayzlin (2006) [13]
Huang, et al. (2015) [14]
Yu, et al. (2012) [15]
Archak, et al. (2011) [16]
Lv (2019) [17]
Jiang and Li (2021) [18]
Zhang, et al. (2022) [19]
Jiang, et al. (2022) [20]
Geva, et al. (2017) [6]
Choi and Varian (2012) [22]
Huang, et al. (2017) [24]
Li, et al. (2017) [25]
Liu, et al. (2011) [26]
Kang, et al. (2020) [27]
Wang, et al. (2015) [28]
Jiang, et al. (2021) [29]
Yao, et al. (2021) [30]
Jin, et al. (2022) [31]
Ouyang, et al. (2020) [36]
Chang (2020) [37]
Liu and Zhang, et al. (2021) [38]
Liu and Shen, et al. (2021) [39]
Kong and Zhu (2018) [40]
Pamuksuz, et al. (2021) [41]
Alantari, et al. (2022) [42]
Zhao (2014) [43]
Chen (2011) [44]
Cui (2014) [45]
Liu, et al. (2017) [46]
Zhang (2020) [47]
Current study
Table 2. Meaning of variables and symbols.
Table 2. Meaning of variables and symbols.
Variables and SymbolsMeaning
y t Sales volume in the tth month
q The effect of the sales data of the first q months on the sales of the tth month
p The effect of online public opinion in the first p months on the sales of the tth month
m The effect of Baidu Index in the first m months on the sales of the tth month
w t t -period online public opinion index
w k t t -period KOL online public opinion index
b t t -period Baidu Index
α i , β j , γ k Model parameters obtained by least squares regression
d t Time dummy variable
ε t Error term
Table 3. Monthly Weibo sentiment index data.
Table 3. Monthly Weibo sentiment index data.
MonthNIOXPeng
January, 20180.6556741030.73715415
February, 20180.5770833330.677419355
March, 20180.6500904160.711805556
April, 20180.5447347590.719
May, 20180.507552870.554179567
June, 20180.6934174930.608540925
July, 20180.5702479340.102443609
August, 20180.4026595740.385777778
September, 20180.4765297570.615023474
October, 20180.2041002280.562334218
November, 20180.0451888490.644149578
December, 20180.1440694310.67345815
January, 20190.1507754160.602597403
February, 20190.3399236910.704856787
March, 20190.582775920.180915609
April, 20190.2169454940.620724346
May, 20190.5346893290.458024691
June, 20190.088551550.450980392
July, 20190.634456595−0.195234708
August, 20190.5453359430.355384615
September, 20190.1969696970.343719572
October, 20190.4521172640.513580247
November, 20190.1350304370.548323471
December, 20190.5270209920.571528752
January, 20200.3237012990.540114613
February, 20200.3303265940.513064133
March, 20200.3074901450.61826484
April, 20200.305295950.29985082
May, 20200.3983797110.605472637
June, 20200.5026281210.601873536
July, 20200.5174825170.766788424
August, 20200.4739165330.598140496
September, 20200.5649152070.622379778
October, 20200.4304123710.615131579
November, 20200.6180832860.421680097
December, 20200.4121037460.35242994
January, 20210.3550761420.380594577
February, 20210.4001896630.484677244
March, 20210.2205268460.35873388
Table 4. Sales volume data.
Table 4. Sales volume data.
MonthNIOXPeng
July, 20181331
August, 20181296
September, 20182079
October, 20182060
November, 20183349
December, 2018269241
January, 20191803599
February, 2019654600
March, 201913561256
April, 201915082200
May, 201910682704
June, 201910922237
July, 201915021515
August, 20192796306
September, 201924782186
October, 20192019505
November, 201915001016
December, 201923631485
January, 20201854630
February, 2020707710
March, 20201624789
April, 202030151008
May, 20203563954
June, 20204018821
July, 20203680551
August, 20203761623
September, 20205003853
October, 20205145815
November, 202055004650
December, 202066236420
January, 202177485180
February, 202158903035
March, 202174494423
Table 5. Some Baidu Index data.
Table 5. Some Baidu Index data.
KeywordAreaDateSearch IndexPC TrendsMobile Trends
NIOChina1 January 201835647692795
NIOChina2 January 2018404113202721
NIOChina3 January 2018421514152800
NIOChina4 January 2018346011332327
NIOChina5 January 2018307110062065
Table 6. Monthly Baidu Index data.
Table 6. Monthly Baidu Index data.
MonthNIOXPeng
January, 2018184,87473,316
February, 201899,25951,380
March, 2018138,02260,253
April, 2018110,54969,368
May, 2018106,93558,378
June, 201890,57749,291
July, 2018113,651126,877
August, 2018115,896106,661
September, 2018154,38651,441
October, 2018174,66565,081
November, 2018214,52267,907
December, 2018205,548164,572
January, 2019143,54491,201
February, 2019123,82458,913
March, 2019161,65187,016
April, 201978,727114,502
May, 201975,780246,014
June, 201997,514175,437
July, 201977,540162,598
August, 2019107,68397,244
September, 2019164,083337,566
October, 2019121,971459,830
November, 201955,232656,090
December, 201957,5871014,506
January, 202049,56190,245
February, 202050,80560,876
March, 202049,709202,587
April, 202047,267192,589
May, 202051,140537,397
June, 202054,071548,768
July, 202072,603490,182
August, 202068,538239,659
September, 202056,639272,793
October, 202068,250154,015
November, 202086,946372,831
December, 202068,321190,401
January, 202192,206232,038
February, 202153,704134,965
March, 202166,905179,304
Table 7. Main parameters of the model.
Table 7. Main parameters of the model.
Parameter NameParameter Value
input_size1
hidden_size4
num_layers2
lr0.01
Table 8. Comparison of the prediction accuracy of the four models.
Table 8. Comparison of the prediction accuracy of the four models.
ModelsMAPE
NIOXPeng
Sales prediction model based on historical sales24.6600837.7518
Sales prediction model based on online public opinion and online search index20.3515729.77872
Sales prediction model based on KOL online public opinion and online search index16.3693418.87666
LSTM sales prediction model based on KOL online public opinion and online search index9.77185.899
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Zhang, M.; Xu, H.; Ma, N.; Pan, X. Intelligent Vehicle Sales Prediction Based on Online Public Opinion and Online Search Index. Sustainability 2022, 14, 10344. https://doi.org/10.3390/su141610344

AMA Style

Zhang M, Xu H, Ma N, Pan X. Intelligent Vehicle Sales Prediction Based on Online Public Opinion and Online Search Index. Sustainability. 2022; 14(16):10344. https://doi.org/10.3390/su141610344

Chicago/Turabian Style

Zhang, Mingyang, Heyan Xu, Ning Ma, and Xinglin Pan. 2022. "Intelligent Vehicle Sales Prediction Based on Online Public Opinion and Online Search Index" Sustainability 14, no. 16: 10344. https://doi.org/10.3390/su141610344

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop