Predicting how a crashing or bubbling asset affects a financial network through sentiment diffusion

金融学中的网络理论帮助建立了不同公司和资产之间的潜在关系模型。这样做的目的是了解如果特定资产的表现高于或低于预期，资产网络将如何反应。我的目标是利用金融新闻和社交媒体数据来生成一个网络，并预测当一种资产进入当前金融文献中定义的“泡沫”或“崩溃”状态时，情绪将如何在该资产网络中扩散。这项研究之所以与金融界相关，是因为它可以揭示哪些资产可能受到正在经历极端市场事件的资产的积极或消极影响。在第一阶段，我将通过微调语言模型来开发一种更准确的方法来确定金融文章的语义极性。我将通过将模型暴露给大型金融新闻语料库来实现这一点，以便它理解金融术语，并且我将利用自然语言处理的最新进展来提供最先进的性能。创建这种语言模型并使其开源，将使其他人更容易实现最先进的性能。接下来，为了生成一个金融网络，我将使用一个共现网络，如果资产出现在相同的新闻文章中，连接会更强，并根据不同资产的说法来比较网络。为此，我将利用Masked Language Model（MLM）训练目标来微调语言模型，该模型将在给定时间段内计算出下一个最有可能取代掩蔽资产名称的资产。通过这样做，我将知道哪些资产以类似的方式被写入。这有可能暴露公司之间的潜在关系。最后，为了分析情绪如何在网络中扩散，我将把这种扩散与流行病学模型进行比较，在流行病学模型中，网络的节点被分为易感、感染和康复类别。为了预测情感如何在网络中传播，我将充分利用网络嵌入和链接预测方面的最新进展。这最后一个阶段的含义是，组织将能够预测泡沫资产将如何影响其网络中的不同资产。

Network theory within finance has helped model the underlying relationships between different companies and assets. The purpose of this is to understand how a network of assets will react if a particular asset performs above or below expectations. It is my goal to use financial news and social media data to generate a network and predict how sentiment will diffuse through that network of assets when one asset enters a "bubbling" or "crashing" state as defined in the current finance literature. The reason why this research is relevant to the financial community is that it could reveal information about which assets are likely to be positively or negatively affected by an asset that is going through an extreme market event.In the first stage, I will develop a more accurate way of determining the semantic polarity of a financial article by fine-tuning a language model. I will do this by exposing the model to a large financial news corpus so that it understands financial terms, and I will utilise recent advancements in Natural Language Processing to deliver state-of-the-art performance. Creating this language model and making it open source will make it easier for others to achieve state-of the-art performance.Following that, to produce a financial network I will use a co-occurrence network where connections are stronger if assets appear in the same news articles and compare that a network based on how different assets are spoken about. To do this I will utilise the Masked Language Model (MLM) training objective to fine-tune a language model which will work out the next most likely assets to take the place of a masked asset name over a given time period. By doing this I will know which assets have been written about in a similar way. This has the potential to expose underlying relationships between companies. Finally, in order to analyse how sentiment will diffuse through a network I will compare the diffusion to an epidemiological model in which nodes of the network are split up into susceptible, infected and recovered categories. In order to predict how sentiment will diffuse through a network I will make the most of recent advancements in network embedding and link prediction. The implication of this final stage is that organisations will be able to predict how a bubbling asset will affect different assets within its network.