Fin-ALICE: Artificial Linguistic Intelligence Causal Econometrics

Abstract

This study introduces Fin-ALICE (Artificial Linguistic Intelligence Causal Econometrics), a framework designed to forecast financial time series by integrating multiple analytical approaches including co-occurrence networks, supply chain analysis, and emotional sentiment analysis to provide a comprehensive understanding of market dynamics. In our co-occurrence analysis, we focus on companies that share the same emotion on the same day, using a much shorter horizon than our previous study of one month. This approach allows us to uncover short-term, emotion-driven correlations that traditional models might overlook. By analyzing these co-occurrence networks, Fin-ALICE identifies hidden connections between companies, sectors, and events. Supply chain analysis within Fin-ALICE will evaluate significant events in commodity-producing countries that impact their ability to supply key resources. This analysis captures the ripple effects of disruptions across industries and regions, offering a more nuanced prediction of market movements. Emotional sentiment analysis, powered by the Fin-Emotion library developed in our prior research, quantifies the emotional undertones in financial news through metrics like “emotion magnitude” and “emotion interaction”. These insights, when integrated with Temporal Convolutional Networks (TCNs), significantly enhance the accuracy of financial forecasts by capturing the emotional drivers of market sentiment. Key contributions of Fin-ALICE include its ability to perform month-by-month company correlation analysis, capturing short-term market fluctuations and seasonal patterns. We compare the performance of TCNs against advanced models such as LLMs and LSTMs, demonstrating that the Fin-ALICE model outperforms these models, particularly in sectors where emotional sentiment and supply chain dynamics are critical. Fin-ALICE provides decision-makers with predictive insights and a deeper understanding of the underlying emotional and supply chain factors that drive market behaviors.

1. Introduction

Financial market analysis has become increasingly complex, requiring the integration of diverse data sources and advanced analytical techniques to navigate the intricate landscape of global markets. The period from 2019 to 2021, characterized by geopolitical shifts, economic sanctions, pandemics, natural disasters, and wars, presented a particularly volatile and challenging environment for financial decision-making (McCarthy and Alaghband 2023a). In this context, the incorporation of emotional sentiment analysis into traditional market analysis has emerged as a promising approach to gain a richer understanding of market dynamics (McCarthy and Alaghband 2023b).

The assumption that sentiment analysis can enhance the capacities of stock market forecasting models has been a driving force behind recent research efforts (Liapis and Karanikola 2023). Numerous studies have explored the application of deep learning techniques, such as Temporal Convolutional Networks (TCNs) (Liapis and Kotsiantis 2023; Guo et al. 2023; Dai et al. 2022; Guo et al. 2022; Cen et al. 2022), and the integration of sentiment analysis (Correia et al. 2022; Chong et al. (2017)) to improve stock market prediction. These approaches have shown promising results, outperforming traditional econometric models (Pothepalli 2021).

Moreover, the rapid advancements in large language models (LLMs) have sparked interest in their potential applications in the financial domain. Models such as FinBERT-LSTM (Halder 2022), LagLlama (Rasul et al. 2024), and FinGPT (Liu et al. 2023; Wang et al. 2023) have been developed to address the unique challenges of financial text data. Recent studies have explored the capacity of LLMs like ChatGPT in forecasting stock price movements (Lopez-Lira et al. 2023), highlighting both the potential and limitations of these models in financial prediction tasks. Despite the impressive capabilities of LLMs in natural language processing tasks, their performance in time series forecasting tasks has lagged behind specialized models like TCNs (Jin et al. 2023; Miller et al. 2024).

Recent studies have also highlighted the importance of considering supply chain dynamics in financial market analysis. McCarthy and Alaghband (2023b) propose a multidimensional data science framework that synthesizes data on world economies, stock metrics, significant market events, and knowledge concepts to construct a knowledge graph that captures the relationships between supply chain disruptions, corporations, and commodities.

Traditional approaches to analyzing company correlations often rely on longer-term averages, which can obscure important short-term dynamics and events that drive market behaviors. In this study, we shift to a month-by-month correlation analysis, to better capture the evolving nature of these relationships and align them with specific economic, political, or industry-specific events. This granular approach not only enhances our understanding of the mechanisms by which news and market movements propagate through the corporate ecosystem but also provides a more responsive and adaptable framework for risk management and portfolio optimization.

Building upon these developments, we introduced a novel quantitative metric called “emotion magnitude” that captures the emotional undercurrents of the market (McCarthy and Alaghband 2023b). By integrating this metric with traditional time series analysis using TCNs applied to stock market futures, we demonstrated a more holistic understanding of market dynamics. Our findings suggest that incorporating emotion magnitude as a feature leads to significantly better performance in predicting future market trends compared to traditional market-based risk measures. This granular approach enhances our understanding of how news and market movements propagate through the corporate ecosystem by focusing on the most recent events and interactions. For example, in the shorter one-month horizon, we identified a significant co-occurrence between Goldman Sachs and Amazon in March 2019. This relationship was not visible in a two-year aggregated view but becomes relevant when considering Goldman Sachs’ digital strategy, which included collaboration with Amazon Web Services (AWS) during this period. Similarly, in January 2020, we observed a correlation between Boeing and Honeywell. This is further validated by a news article in March 2020, where Boeing and Honeywell formally announced a partnership, reflecting the utility of short-term correlation analysis in capturing emerging relationships.

In this paper, we present a comparative study using Fin-ALICE (Artificial Linguistic Intelligence Causal Econometrics), a comprehensive framework for financial time series forecasting that integrates emotional analysis (i.e., artificial linguistic intelligence) of financial news, co-occurrence networks (i.e., causal econometrics), and supply chain analysis. By leveraging a shorter time horizon in our co-occurrence network, we enable more near-term analysis, while incorporating a feature selection process to dynamically identify the best-correlated features across sectors. We focus specifically on how the integration of emotional sentiment and supply chain analysis enhances short-term market trend predictions. Our goal is to demonstrate that Fin-ALICE not only outperforms these models but also offers more granular insights into market behaviors by considering emotional and supply chain factors. The outcomes of our research provide a robust analytical tool for financial risk strategy, empowering stakeholders to navigate the complexities of the ever-evolving global financial ecosystem with greater precision and understanding.

This paper is structured as follows: Section 2 presents related work. Section 4 presents the methods and materials used in our research for the development of the Fin-ALICE framework, including the integration of emotional analysis, supply chain data, and advanced deep learning techniques such as Temporal Convolutional Networks (TCNs) and large language models (LLMs). Section 5 provides a conclusion on the improved performance of the emotion interaction feature and the month-by-month correlation analysis, as well as findings related to the relationships between commodity-producing countries, supply chain events, top companies by sector, and their impact on market dynamics.

2. Related Work (Used for Comparative Study)

The intersection of deep learning and financial market analysis has given rise to innovative approaches that aim to enhance forecasting accuracy by integrating various data sources and techniques. This section reviews the related works that have contributed to the development of models combining emotional sentiment analysis, supply chain dynamics, and advanced deep learning techniques for financial market prediction.

For each related work presented, we detail the methodologies used, their contributions to the field, and how they are implemented or adapted in our study for direct comparison with our proposed model. This comparative analysis provides a clear understanding of where current methodologies excel, their limitations, and how our proposed approach addresses these gaps to advance the state of the art in financial time series forecasting.

2.1. LSTM + BERT

The integration of Long Short-Term Memory (LSTM) networks with Bidirectional Encoder Representations from Transformers (BERT) has been explored to improve the handling of sequential and textual data in financial forecasting. The model FinBERT-LSTM, introduced by Halder (2022), leverages the contextual understanding of BERT to process financial news and the temporal capabilities of LSTM for sequential data. This hybrid approach aims to capture complex patterns in financial time series data, demonstrating enhanced prediction accuracy compared to traditional models. The implementation details and further explorations are available in the code repository (xraptorgg 2024). In our work, we compare our TCN model with the FinBERT-LSTM model and also compare the FinBERT LLM model’s emotional analysis capabilities with our emotion interaction feature. FinBERT LLM (Araci 2019) is a large language model specifically fine-tuned on financial texts, distinct from LSTM in its architecture and ability to process and generate human-like text. FinBERT LLM can understand and generate complex financial language, potentially offering different insights into sentiment analysis.

2.2. Temporal Convolutional Networks (TCNs) + MultiLabel BERT

Temporal Convolutional Networks (TCNs) combined with MultiLabel BERT models have shown promise in extracting nuanced emotional sentiment from financial news to inform market predictions. Guo et al. (2023) describe a methodology where textual data are pre-processed and passed through sentiment and emotion classification modules. The study utilizes various sentiment analysis tools like TextBlob, Vader, and FinBERT to generate sentiment polarity time series, along with 28 distinct emotion time series. A feature selection procedure, employing the SelectKBest class from the scikit-learn library (Pedregosa et al. 2011), refines these inputs for the TCN model. Despite challenges in accessing the original code, the procedure outlined in their study offers valuable insights into the integration of sentiment and deep learning models for financial analysis.

In our study, we replicated this multilabel emotion classification process using 28 emotion categories based on the GoEmotions dataset (Monologg 2024; Demszky et al. 2020), and included all sentiment analysis tools mentioned in the paper. We applied the SelectKBest feature selection and a 7-day rolling mean to smooth the generated time series data. Leveraging our TCN network architecture, we evaluated the effectiveness of each feature in predicting financial trends. Additionally, we introduced our own emotional features, such as emotion interaction, to assess their comparative performance.

This replication enabled us to directly compare the impact of various emotion classification methods on the predictive power of the TCN model. Our results show that while traditional classification tools like TextBlob and Vader provide good baseline performance, the inclusion of multilabel emotion classification and our custom emotional features significantly enhance model accuracy by capturing more complex emotional dynamics within financial news.

2.3. FinGPT Forecaster

The FinGPT Forecaster, as detailed by Liu et al. (2023), explores the potential of fine-tuning large language models (LLMs) for financial forecasting. Although not primarily designed for time series, FinGPT demonstrates the adaptability of LLMs in the financial domain through prompt engineering and fine-tuning. The implementation focuses on the Dow Jones Industrial Average (DJIA) using the Llama2-7B model, which is employed for natural language processing tasks such as sentiment analysis and news interpretation, highlighting the versatility of LLMs in financial applications. The corresponding code repository provides further implementation details and potential modifications for time series analysis (AI4Finance-Foundation 2024). In this research, we give the FinGPT forecaster a 7-day forecasting task, and provide the forecaster with time series information from the previous 30 days to predict the next 7 days. We compare the performance of our TCN model with the FinGPT Forecaster model.

2.4. Time-LLM

Jin et al. (2024) introduce Time-LLM, a novel approach that reprograms large language models for time series forecasting. This methodology leverages the inherent capabilities of LLMs in processing sequential data, adapting them specifically for time series tasks. The model demonstrates significant improvements in forecasting accuracy, providing a robust framework for incorporating LLMs into financial market analysis. The official implementation offers comprehensive details on the model’s architecture and performance (KimMeen 2024). The Time-LLM model depends on the previous five closing prices for each forecast day. In our research, we compare the performance of our TCN model with the Time-LLM model.

2.5. Lag-Llama

The Lag-Llama model, presented by Rasul et al. (2024), addresses the challenges of time series forecasting by integrating large language models with lagged data inputs. This approach emphasizes the importance of temporal dependencies and lag structures in financial data, enhancing the model’s predictive capabilities. The implementation details and further exploration of this model are available in their code repository (time-series-foundation-models 2024). In this research, we give Lag-Llama model a 7-day forecasting task, and provide the forecaster with time series information from the previous 30 days to predict the next 7 days. We compare the performance of our TCN model with this large language model.

2.6. Supply Chain Dynamics and Sentiment Analysis

In McCarthy and Alaghband (2023b), we proposed a multidimensional data science framework that integrates supply chain dynamics with sentiment analysis for long-term financial market prediction over a two-year period. In this study, we build upon that work by adapting the framework for a monthly analysis, introducing a new feature, emotional interaction. In addition, we refine several existing state-of-the-art models and perform a significant comparative study, which demonstrates that our enhanced model outperforms these models in predictive accuracy. In our approach, we synthesize data on world economies, stock metrics, significant market events, and knowledge concepts to construct a knowledge graph. This graph captures the relationships between supply chain disruptions, corporations, and commodities, providing a comprehensive understanding of market dynamics. This work highlights the potential of combining diverse data sources to enhance financial forecasting models.

We implemented a comprehensive data science framework that integrates diverse data sources to construct a knowledge graph for financial market analysis. The first step in this framework was to collect the necessary data, which include the following:

• Macroeconomic Data: Macroeconomic data were sourced from the Federal Reserve Economic Data (FRED), providing insights into economic indicators and trends that influence broader market conditions.
• Sector-specific Financial Market Data: Sector-related financial data were obtained from Yahoo Finance using sector-specific indices and exchange-traded funds (ETFs) for the top 10 companies in each sector. The sector indices used include the following:
  • Consumer Staples: XAP=F.
  • Utilities: XAU=F.
  • Materials: XAB=F.
  • Consumer Discretionary: XLY.
  • Healthcare: XAV=F.
  • Real Estate: XLRE.
  • Energy: XAE=F.
  • Industrials: XAI=F.
  • Financials: XAF=F.
  • Technology: XAK=F.
  • Telecommunications: XAZ=F.
  • Gold: QO=F.
  These indices were used as proxies to represent the overall performance of each sector and capture key market movements that could impact sentiment and company performance.
• Financial News and General News Articles: Financial news articles were collected from a variety of sources, including well-known financial platforms such as Nasdaq, Barrons, TheStreet, Investing.com, Forbes, MarketWatch, and Bloomberg. To complement this, we also included news articles from general and international sources like The New York Times, The Washington Post, Reuters, Fox News, CNN, BBC, and CNBC to provide a broader context of events impacting market sentiment.
• Emotion Analysis in News Data: The news data were processed through our emotion library and various sentiment analysis models, including TextBlob, Vader, FinBERT, and a custom multilabel emotion classification model based on the GoEmotions dataset to support our comparison. This allowed us to extract not only polarity (positive, neutral, negative) but also nuanced emotional categories that could indicate investor sentiment. These enhancements represent an improvement over our previous methodology and are necessary to compare the performance of our new features with those used in state-of-the-art models.
• Knowledge Graph Construction and Feature Integration: Using the collected data, we constructed a knowledge graph that captures the relationships between supply chain disruptions, corporations, and commodities. The framework synthesizes the data to model interactions at multiple levels, including global economic indicators, sector-specific movements, and company-level activities. We incorporated a smaller horizon as a parameter to ensure the knowledge graph captures short-term market dynamics and evolving relationships between companies.

Building on this framework, we introduced a novel quantitative metric called "emotion magnitude" to capture the emotional undercurrents of the market. By integrating emotion magnitude with traditional time series analysis using TCNs, we demonstrated a significant improvement in predicting future market trends. Our findings indicate that incorporating emotional sentiment as a feature can lead to better performance compared to traditional market-based risk measures, particularly in sectors where emotional responses and supply chain dynamics play a critical role.

This combined approach underscores the value of synthesizing emotion metrics with supply chain data to provide a more nuanced understanding of market behaviors. It highlights the synergy between short-term emotional fluctuations and long-term supply chain effects, paving the way for more accurate and robust financial and market-based risk forecasting models.

In the following sections, we will detail our proposed work and methods, present our findings, and discuss their implications for the field of financial forecasting.

3. Proposed Work

The landscape of financial market analysis has evolved significantly, and the reviewed studies underscore the potential of integrating deep learning techniques, sentiment analysis, and supply chain dynamics. Models like FinBERT-LSTM, TCN with MultiLabel BERT, FinGPT Forecaster, Time-LLM, and Lag-Llama showcase advancements in this domain, each contributing unique methodologies and insights. These developments, along with the integration of emotional sentiment analysis, large language models (LLMs), and granular company correlation analysis, have led to more sophisticated forecasting models. Building on the landscape of financial market analysis described above, our study will introduce several key innovations:

• Enhanced Emotional Sentiment Analysis: We expand on our previous work with the “emotion magnitude” metric and introduce a new feature called “emotion interaction”. The emotion interaction feature is derived from the count of companies (how often they show in news articles) and emotion sentiment data. This feature captures the interplay between emotional sentiment and company activity levels, providing a more refined view of market dynamics in a single metric. We integrate these features with Temporal Convolutional Networks (TCNs) with the goal of improving the prediction of future market trends.
• Leveraging Large Language Models (LLMs): We explore the potential of LLMs in financial analysis by comparing their time series forecasting capabilities against temporal models such as LSTM and TCN. We enhanced the FinGPT model by incorporating supply chain analysis along with 30 days of time series data and the emotion magnitude feature, embedding knowledge of sentiment-driven supply chain events. In contrast, the Lag-Llama and Time-LLM architectures did not support the incorporation of additional data beyond the time series itself, limiting their ability to leverage external features. Our study evaluates whether these LLMs can perform as well as temporal models in the financial domain.
• Multidimensional Data Science Framework: We adapted our comprehensive framework that synthesizes data on world economies, stock metrics, and market events to construct a knowledge graph by incorporating a shorter time horizon as a parameter. This modification allows the multidimensional framework to improve the accuracy and depth of financial forecasts compared to existing models, making it more adaptable to rapidly changing market conditions. These findings offer a more dynamic understanding of complex relationships and correlations between different companies and sectors in the economy.

Emotion Interaction Calculation

The emotion interaction is calculated by combining the scaled emotion scores with a scaled version of the company count (how often a company is mentioned). The company count is cubed to give more importance to companies that are frequently mentioned. The equation is defined as follows:

$$E{I}_i=\left(w_e·{\displaystyle \frac{E_i-E_{min}}{E_{max}-E_{min}}}\right)+\left(w_c·{\displaystyle \frac{\left(C_i^3\right)-\left(C_{min}^3\right)}{\left(C_{max}^3\right)-\left(C_{min}^3\right)}}\right)$$

(1)

• $E{I}_i$ is the emotion interaction for the size of fear emotion data within the time series news articles i;
• $E_i$ is the emotion score (count of fear articles) i;
• $C_i$ is the company count (count of a company being mentioned in the news for a day) i;
• $w_e=0.9$ is the weight assigned to the emotion score;
• $w_c=0.1$ is the weight assigned to the company count;
• $E_{min},E_{max}$ are the minimum and maximum values of the emotion scores;
• $C_{min},C_{max}$ are the minimum and maximum values of the company count.

This equation balances two key factors: the emotional sentiment extracted from news data and the frequency with which a company is mentioned (company count). The sentiment is given more importance with a weight of $w_e=0.9$, while the company count has a lower weight of $w_c=0.1$. Cubing the company count ($C_i^3$) emphasizes companies that are mentioned frequently, as they are likely to have a more significant impact on market sentiment. The weights ($w_e=0.9$ and $w_c=0.1$) were selected after extensive experimentation to prioritize the stronger predictive influence of emotional sentiment while incorporating company count as a complementary feature. The cubic transformation of company count ($C^3$) emphasizes high values, effectively capturing the influence of frequently mentioned companies without overwhelming the emotional component, ensuring both features contribute meaningfully to the model’s predictive accuracy. The cubic transformation combined with weighted scaling provided the best results by effectively balancing the two features.

Our study investigated various methods for calculating emotion interaction, including direct multiplication, simple addition, logarithmic scaling, rank-based weighting, categorical weighting, normalized scaling without power transformation, exponential weighting, and emotion and count addition with exponential moving average. These approaches ranged from straightforward calculations to more complex weighted and normalized techniques, each aiming to capture the relationship between emotional sentiment and company mention frequency in financial news. By using the final best-performing equation of weighted combination, the model captures both the intensity of emotions and the prominence of companies in the news, which together provide a more nuanced understanding of investor sentiment and market dynamics.

4. Methods

In this section, we briefly describe the process we used to compare the performance of all the emotion- and sentiment-related features with the month-by-month correlation analysis. We also provide an overview of the data sources, the emotion classification model, LLMs, and the Temporal Convolutional Networks (TCNs) used in our research. We then present the results of our analysis and discuss the implications of our findings.

In this study, we compared various advanced neural network models to evaluate their effectiveness in predicting financial market trends based on a diverse set of features. Figure 1 presents an overview of the different network architectures and feature sets used in our analysis.

4.1. Data Sources and Processing

This study builds upon and enhances our Fin-SupplyChain dataset which integrates financial news data (e.g., Bloomberg, Reuters) from 2019 to 2021, focusing on significant supply chain events such as the U.S.–China trade war, economic sanctions on countries like Iran, natural disasters (e.g., hurricanes, wildfires), pandemics (e.g., COVID-19), and armed conflicts. Data were extracted from financial news articles aligned with commodities such as petroleum, natural gas, gold, and integrated circuits, which were cross-referenced with trade patterns from the Observatory of Economic Complexity (OEC 2023). Emotion annotations, particularly focused on fear, were added using the Fin-Emotion library (FinEmotion 2023).

For this study, we enhanced our dataset by incorporating additional sentiment and emotion analysis techniques necessary for the comparison study as shown in Figure 1. We developed a data processing pipeline that extends our previous work to perform the following steps for each sector:

• Extracts Relevant Information: Identifies supply chain issues (e.g., pandemic, natural disaster), commodities (e.g., integrated circuits), and countries (e.g., Taiwan, China) mentioned in each article using a custom entity extraction function.
• Performs General Sentiment Analysis: Evaluates the overall sentiment polarity of the article using general-purpose tools like TextBlob and VADER (e.g., detecting positive or negative sentiment in a news article related to companies identified) to compare to features in a TCN MultiBERT comparison.
• Conducts Finance-specific Sentiment Analysis: Analyzes sentiments tailored to financial contexts using FinBERT for comparison with FinBERT-LSTM (e.g., identifying positive or negative sentiment in financial articles based on a model trained on financial terms).
• Classifies Emotions: Categorizes the financial news text into multiple emotion categories for a more nuanced analysis using GoEmotions through multilabel emotional classification (e.g., detecting multiple emotions such as “optimism” and “fear” in financial news articles).

This enhanced dataset, which we call ’processed_commodity_sector_news_sentiment’, serves as input for our main analysis.

Table 1. Example Row from the Enhanced Dataset—new features in bold.

Field	Value	Field	Value
Date	1 January 2021	Sector	Financials
Title	News headline	Sentiment	0.2
Emotion	Submission	Supply Chain Issue	Pandemic
Commodity	Financial market	Country	United States
TextBlob Sentiment	0.0053	VADER Sentiment	0.6369
FinBERT Sentiment	0.0	Admiration	0.0
Amusement	0.0	Disapproval	0.0
Disgust	0.0	Embarrassment	0.0
Excitement	0.0	Fear	0.0
Gratitude	0.0	Grief	0.0
Joy	0.0	Love	0.0
Nervousness	0.0	Anger	0.0
Optimism	0.0	Pride	0.0
Realization	0.0	Relief	0.0
Remorse	0.0	Sadness	0.0
Surprise	0.0	Neutral	0.9999
Annoyance	0.0	Approval	0.0
Caring	0.0	Confusion	0.0
Curiosity	0.0	Desire	0.0
Disappointment	0.0

provides an example record from this combined dataset.

We further incorporated the top 10 companies by holding in each sector, analyzing significant market events where a company’s stock decreased by over 2%, to identify the time horizon of the supply chain event (how soon the stock was impacted based on the time from article publication). For the energy sector, the companies included Exxon, Chevron Corporation, Conoco, EOG Resources, Schlumberger, Marathon Petroleum, Pioneer Natural Resources, Phillips 66, Kinder Morgan, and Williams Companies. In the consumer discretionary sector, we analyzed Amazon, Tesla, Inc., The Home Depot, Nike, Inc., McDonald’s, Lowe’s, Starbucks, Target Corporation, Booking Holdings, and TJX Companies. The financial sector included Berkshire Hathaway, JPMorgan Chase, Bank of America, Wells Fargo, Citigroup, Morgan Stanley, Goldman Sachs, BlackRock, Charles Schwab Corporation, and American Express. The sectors and macro indexes analyzed included consumer discretionary, consumer staples, energy, financials, gold, healthcare, industrials, materials, real estate, technology, telecommunications, and utilities.

This enhanced dataset created by our pipeline opened avenues to pinpoint peaks of significant market events, offering a granular view into the sectors that bore the brunt of these supply chain disruptions. It also shed light on specific companies that found themselves impacted by these supply chain events. A unique aspect of our dataset is its ability to capture the emotional tone of articles, facilitating an analysis of emotion distributions over time. To rigorously evaluate our approach, we conducted comprehensive comparisons with state-of-the-art models in financial forecasting described in Section 2.

• We benchmarked against the LSTM+BERT model xraptorgg (2024), which combines the sequential learning capabilities of LSTM with the contextual understanding of BERT.
• We compared our results with our implementation of the TCN+MultiLabel BERT approach Liapis and Kotsiantis (2023), utilizing the following emotion classification model:
  • monologg/bert-base-cased-goemotions-original: A BERT-based model fine-tuned on the GoEmotions dataset, which classifies text into 27 distinct emotional or neutral (for a total of 28) categories using a multilabel approach to capture multiple emotions in a single text.
• We also evaluated our model against cutting-edge language models tailored for financial forecasting:
  • FinGPT Forecaster AI4Finance-Foundation (2024), which leverages the power of large language models for financial prediction.
  • Time-LLM KimMeen (2024), a novel approach that reprograms large language models for time series forecasting.
  • Lag-Llama time-series-foundation-models (2024), a foundation model specifically designed for probabilistic time series forecasting.

Our choice of Temporal Convolutional Networks (TCNs) as the primary model for this study was driven by several key factors. TCNs have demonstrated superior performance in capturing long-range dependencies in time series data, a crucial aspect in financial forecasting where historical patterns can significantly influence future trends. Unlike recurrent neural networks (RNNs) or Long Short-Term Memory (LSTM) networks, TCNs can process inputs in parallel, leading to more efficient training and inference times, especially when compared to the size and time complexity of LLMs, which is particularly useful when dealing with large-scale financial datasets. Additionally, TCNs are highly adaptable and can be easily customized to accommodate different input features and data structures, making them an ideal choice for our multidimensional financial analysis framework leveraging emotion metrics and supply chain data. This comparison highlights the strengths of our TCN-based approach and provides valuable insights into the current limitations and potential future directions for LLMs in financial forecasting tasks.

4.2. Financial Co-Occurrence Graph

To create the financial news co-occurrence network, we modified the analysis horizon from a broader period to a month-by-month basis. Each company pair that shared the same emotion within a given month was tracked and counted. Algorithm M-Graph below describes the monthly co-occurrence graph generation process.

For each company pair sharing emotions in a given month, the weight of the edge was determined by the number of unique emotions shared within that month, rather than the sum of all emotion pairs across two years. Edges were added to the graph if the sum of emotion counts for a company pair was greater than 75% of the maximum sum observed across all pairs. This approach ensures that only the most significant relationships, based on both the variety and frequency of shared emotions, are represented in the graph.

This month-by-month analysis allows for a more granular understanding of how these relationships evolve over time and in response to various market events and economic conditions. It captures seasonal patterns and short-term fluctuations that might be obscured in a longer-term analysis.

In the next section, we present a set of analyses with their corresponding visualizations that will corroborate the insights derived from our comparison of advanced neural network models and insights from the month-by-month analysis of knowledge graphs.

5. Results and Discussion

Our analysis of cross-sector correlations reveals distinct patterns in market behavior, particularly during periods marked by global supply chain disruptions. By implementing Algorithm 1 we significantly enhanced our ability to detect correlated companies. This month-by-month approach, compared to a two-year aggregate view, identifies short-term supply chain events and seasonal patterns that would otherwise be missed in long-term analyses.

Algorithm 1: Monthly co-occurrence graph generation

Input: Grouped monthly data frame (group data into monthly year data slices)

Output: A graph G representing monthly co-occurrences

1: Initialize Graph and Variables

- Create an empty graph G

- Extract unique emotions from df_group_month

- Initialize edge_hash dictionary (keeps track of company pairs with same emotion in same day)

2: Process Emotions and Companies

- For each emotion:

- Group data by emotion

- For each row in the group:

- Generate all combinations of companies

- Update edge_hash with emotion counts for each company pair

3: Add Nodes to Graph

- Add all companies as nodes to G

4: Calculate Edge Weights

- Find minimum and maximum sum of emotion counts

- Sort edges by number of unique emotions

- For each edge:

- Calculate weight as sum of emotion counts

- If weight > 75% of max sum, add edge to G

5: Post-processing

- Remove isolated nodes

- Assign colors to nodes

6: Visualize and Export

- Draw the graph using spring layout

- Export graph in GEXF format

The month-by-month analysis highlighted several periods of particular interest:

• August 2019 (Figure 2): Goldman Sachs and Amazon exhibited a notable correlation that appeared this month that was not identified in the long-term analysis (the news co-occurrence graph from the prior research noted no connections with these companies). This is significant as Goldman Sachs had been discussing their digital strategy, which included collaboration with Amazon Web Services (AWS).
• January 2020 (Figure 3): Boeing and Honeywell showed a notable correlation that was not identified in the long-term analysis. Interestingly, in March 2020, Boeing and Honeywell formally announced a partnership, further validating this observed relationship.
• March 2020 (Figure 4): Microsoft and Morgan Stanley also exhibited a notable connection that was not identified in the long-term analysis. During this period, Morgan Stanley had divested from MSCI but retained a material stake in the company who announced partnering with Microsoft in July 2020. This relationship was further solidified when the two companies announced a strategic partnership in the following year in June 2021.

These findings underscore the importance of monthly analysis in capturing evolving relationships between companies and sectors. By identifying specific connections and tracking their progression, we gain a deeper understanding of how strategic alliances and market conditions shape sector interdependencies over time.

In this section, we present the performance of our models based on the features used for training and validation. The models were trained using historical financial data, sector-specific indices, and emotion-related features derived from news articles. For validation, we used a separate set of financial data that were not included in the training process, ensuring that the models were tested on unseen data to evaluate their predictive accuracy. The differences in features across the tables stem from our exploration of various combinations of features, such as company count, emotion interaction, and added emotional features for comparison, to determine which set of features yields the best results for each sector. The tables present the top-performing model in each sector based on different emotion-related features tested during the experiments.

Analysis of the results presented in

Table 2. Performance comparison of different models and features for the consumer discretionary sector. Rows are arranged in order of best performance (validation MAE) to lowest performance. Feature definitions are as follows: Company count is the number of times a company is mentioned in news articles. Emotion interaction is our feature that balances the emotion and frequency of the company being mentioned in news articles. Emotion represents the count of articles categorized under a specific emotion, such as “fear”. Vader and TextBlob are general sentiment analysis tools used to evaluate the polarity of news articles (positive, neutral, negative). FinBERT is a sentiment analysis tool specifically tailored for financial contexts. The remaining 27 emotion categories, along with the “Neutral” category, are derived from the GoEmotions dataset and capture a wide range of emotional expressions within the text.

Feature	Train MAE	Validation MAE	Model	Complexity # Params
disappointment_smooth	0.0120	0.0103	TCN	<70K
company_count	0.0164	0.0138	TCN	<70K
amusement_smooth	0.0121	0.0148	TCN	<70K
pride	0.0160	0.0182	TCN	<70K
pride_smooth	0.0132	0.0203	TCN	<70K
admiration_smooth	0.0129	0.0220	TCN	<70K
neutral	0.0192	0.0231	TCN	<70K
admiration	0.0132	0.0287	TCN	<70K
fear_smooth	0.0137	0.0340	TCN	<70K
emotion & emotion_interaction	0.0131	0.0541	TCN	<70K
textblob_sentiment	0.0148	0.0617	TCN	<70K
surprise	0.0193	0.0665	TCN	<70K
emotion	0.0124	0.0769	TCN	<70K
Customer Discretionary (no sentiment)	0.1225	0.0892	TimeLLM	Llama-7B
finbert_sentiment	0.0112	0.0969	LSTM	<35K
emotion_interaction	0.0173	0.1146	TCN	<70K
forecast 7 days	-	1.0500	LagLlama	Llama-7B
forecast 7 days	-	1.7900	FinGPT	Llama-7B

,

Table 3. Performance comparison of different models and features for the financial sector. Rows are arranged in order of best performance (validation MAE) to lowest performance. (Features are defined in the Table 2 caption.)

Feature	Train MAE	Validation MAE	Model	Complexity # Params
confusion_smooth	0.0275	0.0230	TCN	<70K
company_count	0.0265	0.0244	TCN	<70K
realization_smooth	0.0319	0.0291	TCN	<70K
approval_smooth	0.0244	0.0304	TCN	<70K
textblob_sentiment	0.0332	0.0328	TCN	<70K
finbert_sentiment	0.0228	0.0331	LSTM	<35K
vader_sentiment	0.0287	0.0342	TCN	<70K
neutral_smooth	0.0256	0.0351	TCN	<70K
emotion_interaction	0.0307	0.0375	TCN	<70K
emotion & emotion_interaction	0.0221	0.0408	TCN	<70K
emotion	0.0243	0.0420	TCN	<70K
Financials (no sentiment)	0.3409	0.3076	TimeLLM	Llama-7B
forecast 7 days	-	4.8300	FinGPT	Llama-7B
forecast 7 days	-	11.5000	LagLlama	Llama-7B

and

Table 4. Performance comparison of different models and features for the energy sector. Rows are arranged in order of best performance (validation MAE) to lowest performance. (Features are defined in the Table 2 caption.)

Feature	Train MAE	Validation MAE	Model	Complexity # Params
finbert_sentiment	0.0246	0.0215	TCN	<70K
emotion_interaction	0.0272	0.0218	TCN	<70K
emotion	0.0290	0.0220	TCN	<70K
emotion & emotion_interaction	0.0184	0.0235	TCN	<70K
approval_smooth	0.0369	0.0266	TCN	<70K
finbert_sentiment	0.0203	0.0301	LSTM	<35K
textblob_sentiment	0.0285	0.0325	TCN	<70K
vader_sentiment	0.0200	0.0327	TCN	<70K
neutral_smooth	0.0230	0.0335	TCN	<70K
finbert_sentiment_smooth	0.0377	0.0531	TCN	<70K
Energy (no sentiment)	0.2200	0.2039	TimeLLM	Llama-7B
forecast 7 days	-	26.8500	LagLlama	Llama-7B
forecast 7 days	-	32.5100	FinGPT	Llama-7B

(presented in order of performance) reveals that our Temporal Convolutional Network (TCN) model demonstrates the best and most robust performance across various sectors when utilizing our proposed features such as company count, emotion interaction, and emotion consistently yield competitive results. In several instances, the TCN model’s performance is further enhanced when leveraging the GoEmotion multiclass emotion analysis of news articles as seen with the disappoint_smooth and confusion_smooth, indicating the value of nuanced sentiment information beyond fear in financial forecasting.

Within the group of large language models (LLMs) we evaluated (FinGPT, TimeLLM, and LagLamma), TimeLLM demonstrated the best performance, which we attribute to its innovative architecture incorporating a reprogramming layer. This layer, implemented as a custom attention mechanism, allows for a dynamic mapping between the input time series data and the LLM’s embedding space. However, our observations align with the findings of Tan et al. (2024), who conducted extensive studies on LLM-based time series forecasting methods. Their research highlighted several key insights relevant to our evaluation:

Tan et al. (2024) found that LLM-based methods often exhibit computational inefficiency, requiring significantly more compute time and resources compared to simpler models like our Temporal Convolutional Network (TCN), without yielding proportional improvements in forecasting performance. This inefficiency limits the practicality of using LLMs for time series forecasting in financial applications, especially in scenarios requiring rapid computation and low latency.

Additionally, their study demonstrated that simpler models, such as temporal architecture (TCNs and LSTMs in our study) with standard attention mechanisms, could achieve comparable or superior performance relative to complex LLM-based forecasters, reinforcing the efficacy of more traditional temporal models.

Our experiments with zero-shot (without seeing examples) applications of larger language models, including LagLlama and FinGPT, showed that these models struggled to effectively leverage the 30 days of time series data provided. This highlights the limitations of general-purpose language models in specialized forecasting tasks without domain-specific fine-tuning or architectural modifications. These findings are consistent with broader criticisms of LLMs’ capabilities in complex domains, as noted by Arkoudas (2023), who argues that current LLMs often struggle with tasks requiring genuine reasoning and domain-specific knowledge.

Similar conclusions were drawn in the FinanceBench study by Islam et al. (2023), which identified significant challenges for LLMs in handling financial queries, particularly those requiring numerical reasoning and time-sensitive analysis. Even advanced models like GPT-4-Turbo showed limitations, often providing incorrect answers or refusing to respond to complex financial questions. While augmentation techniques such as using longer context windows showed some improvement in both studies, they also introduced practical challenges such as increased latency and inability to efficiently handle larger financial documents.

Overall, these findings reinforce the value of using simpler models, such as TCNs, which demonstrated superior performance in capturing nuanced sentiment and emotion features with significantly lower computational overhead.

We measure the performance using the formula for Mean Absolute Error (MAE), which quantifies the average magnitude of errors in predictions, allowing for an intuitive assessment of model performance:

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n|y_i-{\widehat{y}}_i|$$

Here, $y_i$ represents the actual values, ${\widehat{y}}_i$ the predicted values, and n the number of observations. The Train MAE is derived from the model’s performance on the training dataset, measuring its ability to fit the data it was trained on. Similarly, the validation MAE is calculated using the validation dataset, offering an evaluation of the model’s predictive capability on unseen data. This distinction helps to assess the model’s generalization performance and the validity of the results.

We performed a correlation analysis between 63 features—including various sentiment and emotion metrics—and the target sector price for the consumer discretionary sector. The features were ranked by their correlation scores with the target price. As shown in Figure 5, the top-performing features for this sector include admiration_smooth, pride_smooth, and amusement_smooth, which highlight the importance of positive emotions in predicting consumer discretionary trends.

The TCN model consistently outperforms other approaches, with features like “disappointment_smooth” and “company_count” yielding the lowest validation MAE (0.0103 and 0.0138, respectively). This suggests that consumer sentiment, particularly negative emotions, plays a crucial role in the consumer discretionary sector performance. The emotion interaction feature, while not the top performer for this sector, still shows competitive results (validation MAE of 0.1146), indicating its potential to capture complex market dynamics.

In the financial sector, the correlation analysis revealed the strongest features predicting the target sector price. Among the 63 features analyzed, emotion_interaction, emotion, and finbert_sentiment_smooth emerged as the most significant as shown in Figure 6. This emphasizes the critical role of nuanced sentiment analysis and its interaction with emotion in the financial market, where market sentiment can strongly influence performance.

The financial sector shows a preference for more complex emotional features, with “confusion_smooth” and “company_count” yielding the best validation MAEs (0.0230 and 0.0244, respectively). This suggests that market uncertainty and the interconnectedness of financial institutions play crucial roles in this sector’s performance.

For the energy sector, correlation analysis across 63 features identified the most influential predictors of the target sector price. As visualized in Figure 7, emotion, emotion_interaction, finbert_sentiment, and vader_sentiment stand out as the top-performing features. This underscores the role of broad sentiment metrics and emotional interplay in the volatile energy market, reflecting its sensitivity to public perception and external events.

In the energy sector, sentiment-based features dominate the top performances. The “finbert_sentiment” feature with TCN yields the best validation MAE (0.0215), closely followed by “emotion_interaction” (0.0218). This underscores the importance of nuanced sentiment analysis in the volatile energy market, where public perception and global events can significantly impact stock prices. Figure 8 shows details of the three models LSTM, TimeLLM, and TCN for the energy sector. The Mean Absolute Error (MAE) for each model is used to measure accuracy, with lower values indicating better performance. The TCN model achieves a validation MAE of (0.0218), which is significantly better than the LSTM’s MAE of (0.0301) and the TimeLLM’s MAE of (0.2039). The TCN model not only shows a lower error rate but also exhibits more stable and consistent predictions, as observed in the visualizations. In contrast, the TimeLLM model displays higher volatility in its predictions, with larger fluctuations and spikes, indicating a less reliable forecast.

We conducted an empirical analysis of stock price movements during August 2019. The analysis focused on 12 identified company pairs across sectors such as technology, consumer discretionary, financials, and energy. Market events were defined as daily price changes exceeding 2%, and we examined whether these pairs moved in tandem or diverged during these events. The results, summarized in

Table 5. Market event analysis for August 2019. This table highlights the stock price changes and co-movement patterns of identified company pairs.

Company 1	Company 2	Event Date	Change (%)	Change (%)	Moved Together
Intel	Apple	2019-08-02	−1.66	−2.12	Yes
Intel	Microsoft	2019-08-05	−3.51	−3.43	Yes
JPMorganChase	GoldmanSachs	2019-08-05	−2.98	−3.67	Yes
Chevron	Exxon	2019-08-05	−1.65	−2.05	Yes
Amazon	JPMorganChase	2019-08-05	−3.19	−2.98	Yes
Microsoft	Target	2019-08-05	−3.43	−0.90	Yes
Visa	Microsoft	2019-08-05	−4.82	−3.43	Yes
Target	Amazon	2019-08-05	−0.90	−3.19	Yes
GoldmanSachs	Citigroup	2019-08-05	−3.67	−3.59	Yes
Microsoft	Amazon	2019-08-05	−3.43	−3.19	Yes
Intel	Apple	2019-08-05	−3.51	−5.23	Yes
JPMorganChase	GoldmanSachs	2019-08-06	0.78	2.15	Yes
Target	Amazon	2019-08-06	2.45	1.29	Yes
Visa	Microsoft	2019-08-06	2.14	1.88	Yes
GoldmanSachs	Citigroup	2019-08-06	2.15	1.64	Yes
Microsoft	Target	2019-08-06	1.88	2.45	Yes
Amazon	JPMorganChase	2019-08-07	0.31	−2.17	No
JPMorganChase	GoldmanSachs	2019-08-07	−2.17	−0.13	Yes
GoldmanSachs	Citigroup	2019-08-08	0.61	2.46	Yes
Intel	Microsoft	2019-08-08	0.94	2.67	Yes
Microsoft	Amazon	2019-08-08	2.67	2.20	Yes
Visa	Microsoft	2019-08-08	2.61	2.67	Yes
Amazon	JPMorganChase	2019-08-08	2.20	1.69	Yes
Chevron	Exxon	2019-08-08	3.47	2.67	Yes
Target	Amazon	2019-08-08	0.95	2.20	Yes
Intel	Apple	2019-08-08	0.94	2.21	Yes
Microsoft	Target	2019-08-08	2.67	0.95	Yes
Intel	Apple	2019-08-09	−2.52	−0.82	Yes
Intel	Microsoft	2019-08-09	−2.52	−0.85	Yes
Chevron	Exxon	2019-08-09	−0.66	−2.13	Yes
JPMorganChase	GoldmanSachs	2019-08-12	−1.88	−2.60	Yes
GoldmanSachs	Citigroup	2019-08-12	−2.60	−2.74	Yes
Amazon	JPMorganChase	2019-08-13	2.21	1.54	Yes
Intel	Microsoft	2019-08-13	2.72	2.07	Yes
Microsoft	Target	2019-08-13	2.07	2.69	Yes
Microsoft	Amazon	2019-08-13	2.07	2.21	Yes
Visa	Microsoft	2019-08-13	1.29	2.07	Yes
Target	Amazon	2019-08-13	2.69	2.21	Yes
Intel	Apple	2019-08-13	2.72	4.23	Yes
Visa	Microsoft	2019-08-14	−2.86	−3.01	Yes
Target	Amazon	2019-08-14	−2.79	−3.36	Yes
Chevron	Exxon	2019-08-14	−3.80	−4.03	Yes
Amazon	JPMorganChase	2019-08-14	−3.36	−4.15	Yes
Intel	Microsoft	2019-08-14	−2.07	−3.01	Yes
JPMorganChase	GoldmanSachs	2019-08-14	−4.15	−4.19	Yes
Microsoft	Amazon	2019-08-14	−3.01	−3.36	Yes
GoldmanSachs	Citigroup	2019-08-14	−4.19	−5.28	Yes
Intel	Apple	2019-08-14	−2.07	−2.98	Yes
Microsoft	Target	2019-08-14	−3.01	−2.79	Yes
Amazon	JPMorganChase	2019-08-16	0.93	2.40	Yes
Intel	Apple	2019-08-16	1.75	2.36	Yes
JPMorganChase	GoldmanSachs	2019-08-16	2.40	1.65	Yes
GoldmanSachs	Citigroup	2019-08-16	1.65	3.52	Yes
Microsoft	Target	2019-08-19	1.67	2.81	Yes
Target	Amazon	2019-08-19	2.81	1.31	Yes
Microsoft	Target	2019-08-21	1.11	20.43	Yes
Target	Amazon	2019-08-21	20.43	1.23	Yes
Target	Amazon	2019-08-22	3.22	−1.04	No
Microsoft	Target	2019-08-22	−0.73	3.22	No
Intel	Apple	2019-08-23	-3.89	−4.62	Yes
Microsoft	Target	2019-08-23	−3.19	−2.66	Yes
Visa	Microsoft	2019-08-23	−2.70	−3.19	Yes
Intel	Microsoft	2019-08-23	−3.89	−3.19	Yes
Target	Amazon	2019-08-23	−2.66	−3.05	Yes
Chevron	Exxon	2019-08-23	−2.17	−2.99	Yes
Amazon	JPMorganChase	2019-08-23	−3.05	−2.48	Yes
GoldmanSachs	Citigroup	2019-08-23	−3.07	−3.07	Yes
Microsoft	Amazon	2019-08-23	−3.19	−3.05	Yes
JPMorganChase	GoldmanSachs	2019-08-23	−2.48	−3.07	Yes
JPMorganChase	GoldmanSachs	2019-08-29	2.27	2.14	Yes
Intel	Apple	2019-08-29	2.36	1.69	Yes
Amazon	JPMorganChase	2019-08-29	1.26	2.27	Yes
Intel	Microsoft	2019-08-29	2.36	1.89	Yes
GoldmanSachs	Citigroup	2019-08-29	2.14	2.47	Yes

, reveal a high degree of correlation, with 71 out of 74 events (95.95%) showing the companies moving in the same direction. This underscores the interconnected nature of markets, where shared sentiment, often influenced by public perception and global events, drives synchronous behavior across sectors. These findings emphasize the critical role of sentiment analysis in understanding market dynamics in these co-occurring companies.

Across all sectors, we observe that our TCN model, when leveraging the proposed metrics (company count, emotion interaction, and specific emotions), consistently outperforms (based on MAE score) more complex models like TimeLLM, LagLlama, and FinGPT. This is particularly evident in the stark contrast between the TCN’s prediction performance and that of the LLMs in the study.

The superior performance of our approach can be attributed to the following:

• The granularity of our emotion-based features, which capture nuanced market sentiments.
• The ability of TCNs to effectively model temporal dependencies in financial time series.
• The integration of supply chain dynamics and company relationships through our monthly co-occurrence graph analysis.

These findings underscore the effectiveness of Fin-ALICE in capturing complex market dynamics and sentiment patterns, offering a more robust framework for financial forecasting compared to general-purpose language models or traditional market indicators.

6. Conclusions

Fin-ALICE represents a significant advancement in financial market analysis, offering a more nuanced and comprehensive approach to predicting market trends. By leveraging emotional sentiment, supply chain dynamics, and advanced machine learning techniques, it provides a powerful tool for navigating the complexities of global financial markets. Our comprehensive comparison of advanced neural network models, including TCNs, LLMs, and specialized time series models, demonstrated the superior performance of our TCN model, particularly when leveraging emotion-based features, consistently demonstrated robust performance across various sectors. Future work should focus on further refining these techniques, exploring the potential of hybrid models that combine the strengths of different approaches, and developing more robust methods for handling the unique challenges of financial time series data. One promising direction would be the creation of a two-stage model where LLMs are used for advanced feature extraction from textual financial news and reports, while TCNs handle the time series forecasting based on these extracted features along with traditional financial risk indicators. This approach could potentially combine the advanced language understanding capabilities of LLMs with the robust time series modeling of TCNs.