我们被要求抄写论文的introduction部分。根据提供的链接内容，我们需要从论文的introduction部分开始抄写，直到introduction结束。
从提供的文本中，我们可以看到introduction部分从“Synthetic generation of labeled images has successfully been ap-”开始，到“Moreover, statistical realism of the augmented time series dataset is often a requirement for explainability purposes in applications such as finance.”结束。
但是注意，提供的文本中，在introduction之后还有背景（BACKGROUND）等部分。因此，我们需要准确提取introduction部分。

然而，用户上传的链接内容中，introduction部分的文本被截断了，因为我们看到在“Multi-agent simulation provides a means of generating synthetic time series using a bottom-up approach. In multi-agent market simulation, the resultant time series originate from interaction of a large number of market agents (e.g., [4]). Because of the flexibility to change market configuration on the agent level, multi-agent simulation is well-suited for modeling counterfactual scenarios, however, it is notoriously difficult to calibrate to reflect realistic properties of the market [22].”之后，才是我们之前理解的introduction的结尾。

重新审视提供的文本，实际上论文的introduction部分应该从“Synthetic generation of labeled images has successfully been ap-”开始，到“Moreover, statistical realism of the augmented time series dataset is often a requirement for explainability purposes in applications such as finance.”结束。但根据我们看到的文本结构，在introduction之前还有摘要（ABSTRACT）和作者信息等。

因此，我们需要抄写的是introduction部分的正文，即从“Synthetic generation of labeled images...”开始，到“... for explainability purposes in applications such as finance.”结束。

但是，我们注意到用户提供的文本中，在“Multi-agent simulation provides a means...”之前的部分是introduction，而“Multi-agent simulation...”开始已经是下一段，但根据论文结构，这仍然属于introduction。

所以，我们抄写从“Synthetic generation of labeled images has successfully been applied to improve generalization capability, and, therefore, performance of machine learning image detection algorithms. Little work, however, has been done on using synthetic time series to improve generalization capability of machine learning forecasting algorithms. In financial services, for example, there is interest in augmenting historical stock price time series with synthetic rare event data in order to enable price prediction models generalize better on unseen scenarios during the unexpected market shocks due to economic events, global pandemic, flash crashes, etc. . One generally expects models trained on such augmented time series to perform well in practice if synthetic time series is statistically similar to real out-of-sample data as suggested by the “train on synthetic, test on real” (TSTR) framework [8]. Moreover, statistical realism of the augmented time series dataset is often a requirement for explainability purposes in applications such as finance.”这一整段结束。

然而，我们注意到在用户提供的文本中，在“Multi-agent simulation provides a means...”之前，还有一句话：“Multi-agent simulation provides a means of generating synthetic time series using a bottom-up approach. In multi-agent market simulation, the resultant time series originate from interaction of a large number of market agents (e.g., [4]). Because of the flexibility to change market configuration on the agent level, multi-agent simulation is well-suited for modeling counterfactual scenarios, however, it is notoriously difficult to calibrate to reflect realistic properties of the market [22].” 这句话也是在introduction部分。

因此，introduction部分应该包括：

1. Synthetic generation of labeled images has ... finance.
2. Multi-agent simulation provides ... market [22].

但是，在论文的introduction部分，通常还会包括接下来的内容，即提出神经风格迁移（NST）作为解决方案。在提供的文本中，接下来是：“In this paper, we propose to use neural style transfer (NST) to improve statistical realism of synthetic time series, and use it to stylize time series that model counterfactual scenarios for the purpose of data augmentation. NST was initially introduced in [12] to incorporate artistic style into photographs and quickly won over the imaging community. By including the feature correlations of multiple layers, one can obtain a stationary, multi-scale representation of the input image (i.e., the Gram matrix), which captures its texture information but not the global arrangement – which is called style representation. Despite receiving attention by the computer vision community, there is little academic work on NST in the context time series. In [20], NST was used to introduce realistic noise to the seismic shock model. In [19], denoising autoencoder combined with NST was used as a generative technique for realistic daily price time series. The generated time series showed statistical properties similar to historical data; additionally, the visual inspection of generated paths indicated presence of technical patterns that are characteristic of the historical daily asset time series. Both of the aforementioned methods extract time series style representations in the same manner they are extracted for images in [12], which may not be suitable notions of styles for realistic time series since they only contain information about correlations across extracted features. Further related work also includes NST for audio and voice synthesis. This class of NST methods utilize a variety of different techniques for extracting style representations including frequency-domain representations (e.g., VGG-19 or wide-shallow-random networks applied to 2D spectrograms [21]), time-domain representations (e.g., SoundNet encoder [3] or WaveNet decoder [17] applied to raw waveforms), and domain expert knowledge of auditory perception [13]. Unfortunately, since these style representations are specific to audio data, they cannot be directly applied to other domains, such as finance.”

然后，在介绍了相关工作的NST之后，introduction部分通常会对本文的主要贡献进行总结。在提供的文本中，接下来是：“Main Contributions: We develop a novel framework for time series stylization with the goal of constructing realistic and useful synthetic datasets. For that, we propose a style transfer method that combines the underlying trend of one time series with the distributional properties of another. These distributional properties, which we refer to as stylized features, are informed by domain knowledge of time series data and can be computed using sample-based approximations, without requiring the training of dedicated feature extraction techniques like convolutional neural networks (CNNs). Trends are extracted by applying time series decomposition on either the original time series dataset or a synthetic dataset generated from another model. To validate our method, we empirically show on three different datasets that our time series style transfer method is able to achieve competitive performance with state-of-the-art time series generation and augmentation schemes.”

因此，introduction部分应该从“Synthetic generation of labeled images...”开始，到上述主要贡献的总结结束。

但是，我们注意到在用户提供的文本中，在“Main Contributions: ...”之后，还有一句话：“Related Work: ...” 这应该属于introduction部分的相关工作，但通常introduction中会包含对相关工作的简要介绍，然后引出本文的工作。

因此，我们判断introduction部分的结束应该在“Main Contributions: ...”之后，因为接下来是“Related Work: ...”，这通常已经是正文的下一部分（即相关工作）了。

然而，在提供的文本中，“Related Work: ...”是introduction部分的一部分，因为它在introduction小节内。所以，我们应该将整个“Related Work: ...”也包含在introduction部分。

但是，我们观察到在“Main Contributions: ...”之后，文本中有一个明显的“Related Work:”，这通常是新的一段的开始，但它在introduction中。所以，introduction部分应该继续到“Related Work: ...”结束。

然而，在提供的文本中，“Related Work: ...”之后还有一段：“Realistic time series generation has been previously studied in the literature by using the generative adversarial net-works (GANs). ...” 这一段也是相关工作，属于introduction的一部分。

直到“Data augmentation is well established in computer vision tasks due to the simplicity of label-preserving geometric image transformation techniques, but it is still not widely used for time series with some early work being discussed in the literature [15]. ... In the computer vision literature, NST was proposed to disassociate content image from style, and hence generate a dataset with new content-style combinations - by utilizing this idea, NST was used as an augmentation technique to improve quality of machine learning classification algorithms for images [28]. In this paper, we investigate the use of NST for realistic time series generation and subsequent data augmentation.” 这段结束，才是introduction的结尾。

因为接下来就是“2 BACKGROUND”了。

所以，introduction部分包括：

1. Synthetic generation of labeled images ... for explainability purposes in applications such as finance.
2. Multi-agent simulation ... market [22].
3. In this paper, we propose ... other domains, such as finance.
4. Main Contributions: ... augmentation schemes.
5. Related Work: ... (直到data augmentation那一段结束)

但是，我们注意到在用户提供的文本中，在“Main Contributions: ...”之后，是“Related Work: Realistic time series generation ...”开始，直到“In this paper, we investigate ... data augmentation.” 结束。

因此，我们需要抄写introduction部分的所有内容。

然而，由于用户要求只抄写introduction部分，我们按照论文的结构，从“Synthetic generation of labeled images...”开始，到“In this paper, we investigate the use of NST for realistic time series generation and subsequent data augmentation.”结束。

注意：我们抄写时要保持原文的格式，包括换行和缩略词（如[8]等），但不需要抄写作者信息、摘要和关键词等。

但是，我们注意到在用户提供的文本中，introduction部分之前有作者信息和摘要，这些不是introduction，所以我们不抄。

因此，我们开始抄写：

Synthetic generation of labeled images has successfully been applied to improve generalization capability, and, therefore, performance of machine learning image detection algorithms. Little work, however, has been done on using synthetic time series to improve generalization capability of machine learning forecasting algorithms. In financial services, for example, there is interest in augmenting historical stock price time series with synthetic rare event data in order to enable price prediction models generalize better on unseen scenarios during the unexpected market shocks due to economic events, global pandemic, flash crashes, etc. . One generally expects models trained on such augmented time series to perform well in practice if synthetic time series is statistically similar to real out-of-sample data as suggested by the “train on synthetic, test on real” (TSTR) framework [8]. Moreover, statistical realism of the augmented time series dataset is often a requirement for explainability purposes in applications such as finance.

Multi-agent simulation provides a means of generating synthetic time series using a bottom-up approach. In multi-agent market simulation, the resultant time series originate from interaction of a large number of market agents (e.g., [4]). Because of the flexibility to change market configuration on the agent level, multi-agent simulation is well-suited for modeling counterfactual scenarios, however, it is notoriously difficult to calibrate to reflect realistic properties of the market [22].

In this paper, we propose to use neural style transfer (NST) to improve statistical realism of synthetic time series, and use it to stylize time series that model counterfactual scenarios for the purpose of data augmentation. NST was initially introduced in [12] to incorporate artistic style into photographs and quickly won over the imaging community. By including the feature correlations of multiple layers, one can obtain a stationary, multi-scale representation of the input image (i.e., the Gram matrix), which captures its texture information but not the global arrangement – which is called style representation. Despite receiving attention by the computer vision community, there is little academic work on NST in the context time series. In [20], NST was used to introduce realistic noise to the seismic shock model. In [19], denoising autoencoder combined with NST was used as a generative technique for realistic daily price time series. The generated time series showed statistical properties similar to historical data; additionally, the visual inspection of generated paths indicated presence of technical patterns that are characteristic of the historical daily asset time series. Both of the aforementioned methods extract time series style representations in the same manner they are extracted for images in [12], which may not be suitable notions of styles for realistic time series since they only contain information about correlations across extracted features. Further related work also includes NST for audio and voice synthesis. This class of NST methods utilize a variety of different techniques for extracting style representations including frequency-domain representations (e.g., VGG-19 or wide-shallow-random networks applied to 2D spectrograms [21]), time-domain representations (e.g., SoundNet encoder [3] or WaveNet decoder [17] applied to raw waveforms), and domain expert knowledge of auditory perception [13]. Unfortunately, since these style representations are specific to audio data, they cannot be directly applied to other domains, such as finance.

Main Contributions: We develop a novel framework for time series stylization with the goal of constructing realistic and useful synthetic datasets. For that, we propose a style transfer method that combines the underlying trend of one time series with the distributional properties of another. These distributional properties, which we refer to as stylized features, are informed by domain knowledge of time series data and can be computed using sample-based approximations, without requiring the training of dedicated feature extraction techniques like convolutional neural networks (CNNs). Trends are extracted by applying time series decomposition on either the original time series dataset or a synthetic dataset generated from another model. To validate our method, we empirically show on three different datasets that our time series style transfer method is able to achieve competitive performance with state-of-the-art time series generation and augmentation schemes.

Related Work: Realistic time series generation has been previously studied in the literature by using the generative adversarial networks (GANs). With the TimeGAN architecture [27], realistic generation of temporal patterns was achieved by jointly optimizing with both supervised and adversarial objectives to learn an embedding space. QuantGAN [26] consists of a generator and discriminator functions represented by temporal convolutional networks, which allows it to synthesize long-range dependencies such as the presence of volatility clusters that are characteristic of financial time series. TimeVAE [6] was recently proposed as a variational autoencoder alternative to GAN-based time-series generation. In [1], Fourier flows, a flow-based generative model for time-series in the Fourier domain, is presented. Fourier flows are able to achieve competitive performance with state-of-the-art generation techniques at much lower training times. Methods like GANs, VAEs and Fourier flows are typically used for learning the underlying distribution of the training data, and may not provide the distributionally new scenarios needed for data augmentation.

Data augmentation is well established in computer vision tasks due to the simplicity of label-preserving geometric image transformation techniques, but it is still not widely used for time series with some early work being discussed in the literature [15]. For example, simple augmentation techniques applied to financial price time series such as adding noise or time warping were shown to improve the quality of next day price prediction model [9], however, such transformations were not required to produce realistic synthetic time series. In the computer vision literature, NST was proposed to disassociate content image from style, and hence generate a dataset with new content-style combinations - by utilizing this idea, NST was used as an augmentation technique to improve quality of machine learning classification algorithms for images [28]. In this paper, we investigate the