OpenSea is the largest marketplace for NFT trading in the world, and according to statistics, it has amassed over $32B in trading volume and over 2.1M traders [8]. NFTs on OpenSea are mainly based on four blockchains, namely Ethereum, Klaytn, Polygon, and Solana, and thus far, more than 95% of the transactions are performed using the Ethereum blockchain. Further, there are many kinds of NFTs on OpenSea, including images, audio and video. The NFT collections can be divided into nine categories: Art, Domain Names, Collectible, Music, Sports, Photography, Trading Cards, Virtual Worlds, and Utility. Each NFT usually has three kinds of related pages, and the information on these pages can be summarized as: NFT information, collection information, and related account information. These three website pages of the NFTs are presented in Fig. 1 to illustrate OpenSea.

The existing studies on NFT price fluctuations are primarily focused on specific sought-after collections. Dowling found that the pricing sequence of virtual lands in Decentraland is characterized by inefficiencies and appreciation in value [4]. In another study, Dowling studied Cryptopunks, Decentraland, and Axie Infinity and found a limited volatility transmission effect between the pricing of these NFTs and Bitcoin prices [12]. From other perspectives, Aharon and Demir used a time-varying parametric vector autoregressive approach to analyze the correlation between the NFT market and financial assets such as stocks [13]. However, we cannot use the sales trends of these well-known collections as a basis for studying the NFT market. That’s because these sought-after collections, like Decentraland virtual lands and Cryptopunks, aren’t representative of the general market.

To date, no research has used machine learning approach to predict NFT sales price fluctuations, and no reports are available on the demonstration of a specific NFT sale-price prediction error, except that of Nadini et al., who studied the predictability of NFT sales using a linear regression model. Subsequently, they used the AdaBoost algorithm to predict whether an NFT will be traded multiple times. However, they did not report the prediction error and used variables that are difficult to access by most investors, such as centrality of buyers and sellers in the network of NFT transactions, and so on [9].

We performed an empirical study with thousands of common NFTs as subjects to generalize the results. We attempted to predict NFT sale prices and fluctuations accurately by using publicly available and easily accessible data as independent variables of our machine learning models. We expect that this study will help the investors understand the NFT sale prices and their fluctuations more quickly and make appropriate investments.

Machine learning can be divided into supervised learning, unsupervised learning, and reinforcement learning [16], and the machine learning algorithms are usually used for clustering, classification, and regression. Supervised learning is the prediction based on the input data, resulting in common decision rules [17]. Machine learning has been used in many areas, including demand forecasting, price prediction, and illness diagnosis [15,18,19].

Due to the changes in OpenSea policies, the application programming interface provided by the website has some limitations. Therefore, instead of using a web crawler to collect the data, we employed an automated testing tool to simulate the behavior of the people visiting the webpages (by using a Python program) and to collect the visited Hyper Text Markup Language links. Finally, XPath analysis was performed to collect data and obtain key information. Our publicly available Python program can be visited using the following uniform resource locator: https://github.com/wzx333222111/ZIXIONG-WANG. We collected 15,000 pieces of NFT information with sale records on the activity page between July 4 and July 7, 2022. We selected the ether chain NFTs for the data collection because of two reasons: (1) to unify the unit of value, and (2) because ether chain is the blockchain used by most NFTs. Various variables expected to be related to the target variables are shown in Table 1.

Next, we preprocessed the collected data, i.e., performed reduction, cleaning, transformation, and normalization of the collected data to improve the prediction accuracy of the model.

Among the 15,000 collected data, we finally selected 8586 datapoints and split them into a training set (7634 data) and a testing set (952 data) for predicting NFT sale prices. The standard deviation and mean values of the variables are presented in Table 2. In addition, we filtered the NFTs with multiple sale price records, finally selected 6136 datapoints, and divided them into training (5358 data) and testing sets (778 data) to predict the fluctuations of the NFT sale prices.

Then, we used regressor and classifier to develop predictive models. We used Orange (Ver. 3.32.0) as the data analysis tool, which integrates many machine learning algorithms, such as RF, Tree, and Gradient Boosting. However, after comparing the preliminary prediction results of these algorithms available in Orange, we selected the RF and AdaBoost algorithms for this study. The schematic of this tool’s operation is depicted in Fig. 5. To improve the reliability of the regression model, we used K-fold cross-validation, which was performed 10 times. Cross-validation is a method in which one model validation technique, among a large number of similar ones, is used to assess how the results of statistical analyses can be generalized to independent datasets [31,32]. In prediction problems, the model is usually provided a known dataset for training (training set) and an unknown dataset to test the model (testing set) [33]. The purpose of cross-validation is to reduce selection bias and overfitting problems [34].

We developed four phases to estimate the predictive performance of NFT information, collection information and related account information in the regression model. Table 3 shows the variables included in each phase. In Phase 1, we input NFT information as predictors. We expect that using only NFT information as predictors yields high prediction accuracy because these variables provide the most basic information about NFT. The related account information and collection information are added to the variables considered in Phase 2 and Phase 3, respectively. We expect that the addition of these variables should lead to an improvement in prediction accuracy, albeit with a much smaller overall contribution than that of the variables in the first phase. Phase 4 includes all the three types of information, and we expect to get the best prediction results in this phase.

In this study, we used root mean square error (RMSE), mean absolute error (MAE) and R-square (R²) to evaluate the performance of the regression models [15,24,35].

We performed a step-by-step analysis from Phase 1 to 4. Phase 1 includes only the NFT information variables, and RF performs better in this phase of the prediction (RMSE = 0.119, MAE = 0.042, and R² = 0.990) (see Table 4). In Phase 2 with account information, we found that AdaBoost achieved better prediction results, and RMSE, MAE and R² (RMSE = 0.047, MAE = 0.007, and R² = 0.998) (see Table 5) showed significant improvements over their Phase 1 counterparts. The prediction results of Phase 3, which combined NFT and collection information, were slightly more accurate than those of Phase 1 (RMSE = 0.075, MAE = 0.009, and R² = 0.996) (see Table 6). However, compared with Phase 2, the prediction error in this phase became larger. Finally, we input all the variables into the prediction model and observed that the results of Phase 4 were better than those of Phase 1 and 3, whereas the RMSE value was slightly larger than those of Phase 2 (RMSE = 0.053, MAE = 0.007, and R² = 0.998) (see Table 7). These results show that the relevant account information contains information directly related to the NFT sale price.

In the classification model, we evaluated the model using F1-score, precision, and recall [36]. We also performed the receiver operating curve (ROC) analysis. The ROC is established by graphing the relationship between the true positive and the false positive rates under different settings of the threshold. The false positives can also be called as false alarm probabilities [37]. The results of the training set are shown in Table 8. The RF classification model exhibited a peak area under the curve (AUC) value of 0.945, whereas the AdaBoost model showed a peak AUC value of 0.853. In addition, considering the precision, recall, and F1 scores, the RF algorithm outperformed the AdaBoost algorithm in determining the fluctuations of NFT sale prices. The ROC analysis results are shown in Table 9.

For further evaluation, we selected 778 datapoints as the testing set, and again predicted the fluctuations of NFT sale prices using the RF model. Table 10 shows results of the testing set. Among the 778 datapoints in the testing set, 461 and 317 points exhibited actual sale price fall and rise, respectively. The RF classification model correctly predicted the price volatility for 700 samples and incorrectly predicted price rise and fall for 36 and 42 samples, respectively. The results of the testing set are shown in Table 11.

After conducting a learning process to determine the fluctuations of NFT sale prices, we further analyzed the variables collected from OpenSea to identify the ones that contained more information for determining the fluctuations of NFT sale prices based on the RF model. Fig. 6 shows the contribution levels of the variables that affected the predictions. NFT all time avg. price (0.025) was the most influential variable for predicting price rise or fall, followed by the number of NFT favorites (0.022), number of creator’s favorites (0.020), creator’s activity status (0.020), and number of owner’s collection (0.011). Among them, NFT favorite means that other users browsed this NFT and clicked the favorite button, which to some extent reflects the recognition and liking of other users for this NFT. Creator’s favorites indicate those NFTs for which the creator of the NFT clicked the favorite button. Creator’s activity indicates whether there has been recent NFT transfers or transaction activities. Both these variables reflect the activity level of the NFT creator. The owner’s collection represents the number of NFTs owned by the owner.

The results of this study have important theoretical implications. First, we determined whether publicly available information on OpenSea helps to predict the price fluctuations of NFTs. According to previous studies, specific famous NFT prices are associated with cryptocurrencies and other financial assets [12,13]. However, these studies were performed without considering the thousands of ordinary NFTs sold publicly in the NFT market. We identified a significant correlation between the sale price fluctuations of these ordinary NFTs and the information publicly available on OpenSea, and herein, suggest that this information can be used to predict the fluctuations of NFT prices.

Second, NFT is a digital asset with considerable public interest, but few researchers have explored the price-prediction approach. Although Nadini et al. obtained the NFT trading information from Ethereum and WAX blockchains and analyzed the predictability of NFT sales using a linear regression model, they used some variables that are difficult to obtain by most investors, such as centrality of buyers and sellers in the network of NFT transactions and did not report their prediction error [9]. In the present study, we extended the study of Nadini et al. by using Python programs to collect NFT information that is readily available to most investors and used this collected NFT information as independent variables. Further, we demonstrated the prediction error by empirical analysis (RMSE = 0.047, and MAE = 0.007).

Third, in addition to the NFT information of interest, such as NFT price history and numbers of favorited and most viewed NFTs, we also evaluated the prediction value provided by the collection and related account information. The results of the classification model for predicting the fluctuations in the NFT sale prices demonstrated the importance of the relevant account information and revealed that the number of creator's favorites, activity status, and number of owner’s collections primarily contribute to the prediction results. Collectively, the results of the regression and classification models indicate that these three features are crucial for predicting the NFT prices, and thus, future studies on NFT price fluctuations should be conducted by considering these three features.

In this study, publicly available data and simple machine learning algorithms were used. As a result, the methodology and results of this study can be easily replicated by academics and practitioners aiming to predict future NFT price fluctuations.

The results of this study also have practical implications: First, we identified some important variables on OpenSea to predict price fluctuations of NFTs. The investors can use these NFT and account information to understand the NFT prices more simply and make appropriate investments. More specifically, we demonstrated that the prediction value of the relevant account information is higher than that of the collection information. Thus, investors can rapidly and accurately predict the NFT prices. Further, using only the NFT and related account information can prevent unnecessary data collection and reduce the calculation time. We suggest that investors can focus on the number of creators’ collections, status of their activity, and number of owners’ collections, because the NFTs created by more active creators have a greater potential. Thus, in simple words, investors need to find such active creators.

Second, NFT practitioners can produce more attractive NFTs and promote them to ensure more people click the favorite button. In addition, practitioners should pay attention to the role of account activity frequency, because the NFTs created by more active accounts will attract more attention and likes. Therefore, it is necessary to increase the number of favorites given to other NFTs as well as increase the NFT trading activities.

Third, we compared the accuracy of AdaBoost and RF, in predicting the NFT prices and their fluctuations and observed that the practitioners can use the AdaBoost and RF algorithms for predicting the NFT prices and price fluctuations, respectively.