FinXABSA: Explainable Finance
through Aspect-Based Sentiment Analysis

Using a ‘keyword hopping’ framework, we began with the keywords ‘nasdaq stock market’ to gather tweets from Q4 2022, leveraging NASDAQ’s prominence with over 3,300 company listings. From approximately 11k tweets, we identified high-frequency keywords (above 100) [27], counting each word once per tweet. After filtering out overly specific or non-financial terms like tesla or cnbc, we added relevant financial keywords such as ‘sharemarket’, ‘stockstobuy’ [28]. The comprehensive keyword list is detailed in the next section.

Tweets were collected via the Twitter API v2 with academic access from Q4 2022. Due to the sheer number of tweets from the Twitter full archive search, we collected the tweets from only the turn of the hour for every hour each day. We utilised the following keywords: stock market, Nasdaq, inflation, investors, friday sharemarket, monday sharemarket, china stock, china market, china economy, recession, Tuesday sharemarket, stock fall, thursday sharemarket, stock market, market rally, wednesday sharemarket, finance, economy, market closes, stock closes, financial market, sharemarket, stockstobuy, sharemarket drops, pandemic stock. In our keyword list, commas mean ‘OR’ and spaces mean ‘AND’ for Twitter API queries. We excluded retweets and limited tweets to English, collecting about 120k tweets for sentiment analysis.

We collected closing stock prices for six companies from Q4 2022, sourced from Yahoo Finance. Focusing on the rising significance of sustainable finance, we analyzed stocks from the sustainable energy sector, contrasting them with traditional energy. Traditional energy stocks are British Petroleum, Exxon, Shell, and sustainable ones are NextEra, Clearway, Brookfield Renewable. Each selected stock holds a significant market share. Stock prices of companies reflect their financial health and are watched by investors and analysts. Analyzing this data reveals market trends and company performance. Fluctuations offer insights into financial stability, growth, and market position, aiding investors and experts in decision-making. In essence, the data on closing stock prices is vital for gauging the market and the companies’ standings.

As explained later, a list of aspects (attributes or components of a sentence) is necessary for our task of ABSA. When working in the context of FSA, these aspects comprise lists of words used daily in the financial world, for example, “share”, “profit” or “risk”. An extensive list of aspects requires a large compilation of text data. Therefore, we exploit previous research in the FSA domain and draw upon the groundwork of [29], [30] and [31]. The justification for making use of these existing aspect lists is that of trusted statistical methods to generate these words, such as Non-negative Matrix Factorisation, Latent Dirichlet Allocation, and Principal Component Analysis, and annotations by experts in the field, with [29] sharing a similar goal to this paper.

Additionally, as these words contain significant financial meaning, utilising them for ABSA enhances explainability. While keeping the words derived from the aforementioned financial literature the overwhelming bulk of our aspect list, we add another 24 financially important keywords from ‘Keyword Hopping’ (III-A). Altogether, we assemble a list of 131 aspects to facilitate our analysis. Out of these, we focus on the 20 aspects that occurred most frequently in the text data (Table I).

Though [38] has considered separate thresholds for r (i.e. $0.4<|r|<0.6$), this paper will consider correlations of $|r|>0.4$ to be statistically significant. For absolute aspect sentiment scores, $y_{p,NEE}$ yields the greatest magnitude for correlation compared to other sustainable energy stocks. Specifically, $y_{p,NEE}$ yields $r$ values of $-0.731,-0.505,-0.471$ with ${x_{fp}}$ corresponding to the (inflation, economy, finance) aspects respectively. Conversely, $y_{p,NEE}$ yields $r$ values of $-0.726,-0.695,-0.55$ & $-0.457$ with ${x_{fn}}$ corresponding to the respective aspects of (economy, inflation, tax, price). ${r}$ values for normalised aspect sentiment scores, ${x_{nfp}}$ & ${x_{nfn}}$, are noticeably smaller in magnitude compared to absolute aspect sentiment scores. $y_{p,NEE}$ yields an ${r}$ value of $0.44$ with ${x_{nfp}}$ for the stockmarket aspect, and an $r$ value of $-0.449$ with ${x_{nfn}}$ for the economy aspect. Additionally, $y_{p,CWEN}$ yields a ${r}$ value of $-0.405$ with ${x_{nfn}}$ for the financial aspect.

We conduct the Granger causality test to determine whether the different aspect sentiment scores, $\boldsymbol{x}$, Granger cause the various stock prices $\boldsymbol{y_{p}}$ at the 0.05 significance level. From table II, we observe that more aspect sentiment scores Granger cause $y_{p,CWEN}$ & $y_{p,NEE}$ compared to $y_{p,BEPC}$. Additionally, we observe that aspect sentiment scores pertaining to aspects (finance, tax) Granger cause sustainable energy stock prices for the most occurrences compared to sentiment scores for other aspects.

We highlight the highest uncertainty coefficient values between respective aspect sentiment scores and traditional energy stock prices. Specifically, from figure 8, we observe that between ${x_{fp}}$ corresponding to the inflation aspect & ($y_{p,BEPC}$, $y_{p,NEE}$, $y_{p,CWEN}$), uncertainty coefficient values are relatively high at 0.159, 0.127 & 0.125 respectively. Relative to other aspects, uncertainty coefficient also has significant value, at 0.133 between ${x_{fp}}$ corresponding to the aspect of bank & $y_{p,BEPC}$. Between ${x_{fp}}$ corresponding to interest & $y_{p,CWEN}$, it also yields a uncertainty coefficient score of 0.111. Among all uncertainty coefficient corresponding to ${x_{fn}}$ & sustainable energy stocks ${\boldsymbol{y_{p}}}$, uncertainty coefficient is highest between ${x_{fn}}$ corresponding to (inflation, cost, report, tax) & $y_{p,BEPC}$, yielding values of 0.166, 0.143, 0.125 & 0.12 respectively. The next most significant is ${x_{fn}}$ corresponding to (inflation, economy) & $y_{p,NEE}$, where uncertainty coefficient values are at 0.119 & 0.116 respectively.

For uncertainty coefficient corresponding to ${x_{nfp}}$ & sustainable energy stocks ${\boldsymbol{y_{p}}}$, the uncertainty coefficient value is highest between ${x_{nfp}}$ corresponding to the aspect financial and $y_{p,BEPC}$, yielding a value of 0.119. Besides this, uncertainty coefficient values are also relatively significant between economy ${x_{nfp}}$ & $y_{p,CWEN}$, yielding a value of 0.079. Finally, uncertainty coefficient yields a value of 0.079 between market ${x_{nfp}}$ & $y_{p,NEE}$. Lastly, for uncertainty coefficients derived from ${x_{nfn}}$ & sustainable energy stocks ${\boldsymbol{y_{p}}}$, the highest values are yielded between ${x_{nfn}}$ corresponding to (financial, bitcoin) & $y_{p,BEPC}$ at 0.138 & 0.112 respectively.

Of the traditional energy stocks analysed, British Petroleum and Shell consistently yield the greatest magnitude for Pearson correlation values.

For absolute aspect sentiment scores, ($y_{p,BP}$, $y_{p,SHEL}$ $y_{p,XOM}$) have $r$ values of $-0.716$, $-0.671$ & $-0.427$ with ${x_{fp}}$ corresponding to the inflation aspect. $y_{p,BP}$ also has r values of $-0.553$, $-0.542$, $-0.462$ & $-0.422$ with ${x_{fn}}$ corresponding to the (economy, inflation, tax, financial) aspects respectively. Additionally, $y_{p,XOM}$ has an r value of $-0.432$ with ${x_{fn}}$ corresponding to the financial aspect. $y_{p,SHEL}$ has r values of $-0.51$, $-0.508$ & $-0.445$ with ${x_{fn}}$ corresponding to the (economy, inflation, tax) aspects, respectively. For normalised aspect sentiment scores, ($y_{p,BP}$, $y_{p,XOM}$ & $y_{p,SHEL}$) have $r$ values of $0.442,0.416,0.408$ respectively with ${x_{nfn}}$ corresponding to the stock aspect.

A greater number of aspect sentiment scores Granger cause traditional energy stock prices relative to sustainable energy stocks. Additionally, as highlighted in table III, more aspect sentiment scores Granger cause $y_{p,BP}$ & $y_{p,XOM}$ compared to $y_{p,SHEL}$. Moreover, aspect sentiment scores pertaining to the financial aspect Granger cause traditional energy stock prices most frequently compared to sentiment scores corresponding to other aspects.

We highlight the greatest uncertainty coefficient values between sentiment scores and stock prices for traditional energy stocks. Specifically, among all uncertainty coefficient values corresponding to ${x_{fp}}$ & ${\boldsymbol{y_{p}}}$, uncertainty coefficient values are greatest between ${x_{fp}}$ corresponding to inflation & $y_{p,BP}$, at 0.261. Uncertainty coefficient values are also relatively high between ${x_{fp}}$ corresponding to stock and ($y_{p,BP}$ & $y_{p,XOM}$), at 0.158 & 0.145, respectively. Among all uncertainty coefficient values corresponding to ${x_{fn}}$ & ${\boldsymbol{y_{p}}}$, the uncertainty coefficient values between ${x_{fn}}$ corresponding to (economy, inflation, china, financial) & $y_{p,BP}$ are highest at 0.29, 0.196, 0.162 & 0.14 respectively. Additionally, the uncertainty coefficient value between ${x_{fn}}$ corresponding to inflation & $y_{p,SHEL}$ is relatively high at 0.14. These uncertainty coefficient values are significant compared to those yielded between ${x_{fn}}$ corresponding to different aspects & ${\boldsymbol{y_{p}}}$ of different traditional energy stocks.

Between ${x_{nfp}}$ corresponding to (bitcoin, stockmarket, tax) & $y_{p,BP}$, uncertainty coefficient values are at 0.089, 0.085 & 0.075 respectively. Furthermore, ${x_{nfp}}$ for price & $y_{p,XOM}$ yields an uncertainty coefficient value of 0.074. These values are significant relative to uncertainty coefficents derived between ${x_{nfp}}$ of other aspects and different traditional energy ${\boldsymbol{y_{p}}}$. Between ${x_{nfn}}$ of (interest, economy) & $y_{p,BP}$, uncertainty coefficient values are at 0.131 & 0.108 respectively. Moreover, the uncertainty coefficient between stockmarket ${x_{nfn}}$ & $y_{p,SHEL}$ is at 0.113. These values are significant relative to uncertainty coefficients derived between ${x_{nfn}}$ of other aspects and various traditional energy ${\boldsymbol{y_{p}}}$.

We will interpret financially meaningful aspects that have a clear correlation, $|r|>0.4$, between different aspect sentiment scores, $\boldsymbol{x}$, and stock prices $\boldsymbol{y_{p}}$. For the aspect tax, ${x_{fn}}$ yields ${-r}$ with $y_{p,NEE}$, $y_{p,SHEL}$ & $y_{p,BP}$. For price, ${x_{fn}}$ yields ${-r}$ with $y_{p,NEE}$. For financial, ${x_{fn}}$ yields ${-r}$ with ${y_{p,XOM}}$ & ${y_{p,BP}}$, while ${x_{nfn}}$ yields ${-r}$ with $y_{p,CWEN}$. For economy, ${x_{fn}}$ yields $-r$ with the $y_{p,NEE}$, $y_{p,SHEL}$ & $y_{p,BP}$, while ${x_{nfn}}$ yields significant $-r$ with $y_{p,NEE}$. In other words, a reduction in negative sentiment scores with respect to the aspects tax, price, financial, economy is correlated with an increase in prices for separate stocks. The economic interpretation of this is that higher stock prices are correlated with a reduction in pessimism (which essentially translates to more optimism) pertaining to the different components for economic and business conditions (taxes, prices, financial situation) as well as the economy in general. Additionally, for stockmarket, ${x_{nfp}}$ yields ${+r}$ with $y_{p,NEE}$. This indicates that greater positive sentiment about the stockmarket is correlated with a rise in nextera stock price.

Our results are consistent with the findings of [39], [40] & [41] which state that positive sentiments usually lead to rises in stock prices. However, our findings also add a greater level of granularity to this proven relationship, as through ABSA, we distinguished between the sentiment relating to different financial aspects. As such, these results possess the capacity to be made explainable and intelligible through economic and financial theory. To elaborate, our findings corroborate with [42], which describes how positive sentiment for economic and business conditions (i.e. tax, prices, financial situation), as well as the economy & stockmarket, can pertain to higher confidence amongst investors, delineating to increases in stock prices.

However, there are also correlation results whose interpretations are less clear. For instance, both ${x_{fp}}$ and ${x_{fn}}$ corresponding to inflation yields $-r$ with $y_{p,NEE}$, $y_{p,SHEL}$ & $y_{p,BP}$. As they have similar r values, these stocks might be more strongly correlated with the frequency of occurrence of the inflation aspect as opposed to sentiment scores. Other aspects such as finance & stock also yield unexpected results. For finance, ${-r}$ is yielded between ${x_{fp}}$ & $y_{p,NEE}$, while for the stock aspect, $y_{p,SHEL}$, $y_{p,BP}$ & $y_{p,XOM}$ yield ${+r}$ with ${x_{nfn}}$. Pearson correlation shows a stronger link between aspect sentiment scores and traditional energy stocks than sustainable ones. Literature suggests Twitter sentiment has limited influence on renewable energy stock prices [43]. We plan to explore if this is because the sustainable energy sector is relatively nascent and its stocks may not align as closely with financial sentiment as traditional energy stocks.

We reference Granger causality results to complement the Pearson correlation results we obtained. In addition to deriving explainable and significant correlations between various financial aspect sentiment scores and stock prices, Granger causality is indicative of the forecasting power of these sentiment scores for stock prices and confirms the interdependent relationship between them. We believe a more robust and useful analyses of correlation can be done by utilising Granger causality in tandem with Pearson correlation. Chiefly, we highlight specific aspect sentiment scores that are not only strongly correlated with stock prices, but also contain sufficient information to forecast future stock prices, highlighting an interdependent relationship. ${x_{fn}}$ corresponding to the financial aspect Granger causes $y_{p,BP}$ &$y_{p,XOM}$, and an explainable and significant correlation is also present between them. On the other hand, ${x_{nfn}}$ for the stock aspect Granger causes $y_{p,BP}$ and shows significant correlation as well.

Studies have shown that the relationship between textual sentiment and stock pricing is intricate [44], and stock pricing is also driven by numerous (sentiment) features [45]. To clarify this complexity, we propose integrating ABSA with the uncertainty coefficient. We enhance interpretability by computing uncertainty coefficient. This variable measures the degree of information provided by aspect sentiment scores with regard to stock prices. As such, we aim to identify the most statistically significant aspect sentiment scores related to stock prices. Our findings indicate that uncertainty coefficient values are higher between absolute aspect sentiment scores and stock prices as compared to normalised sentiment scores, suggesting absolute sentiment scores offer more insight into stock prices. Additionally, uncertainty coefficient values are higher for traditional energy stocks compared to sustainable ones.

Next, we analyze the uncertainty coefficient values to assess which aspect sentiment scores offer the most valuable information regarding stock prices. Among all aspect sentiment scores and stock prices, the highest uncertainty coefficient is observed between ${x_{fn}}$ related to the economy and $y_{p,BP}$. Following closely is the uncertainty coefficient between ${x_{fp}}$ for inflation and $y_{p,BP}$. Furthermore, for $y_{p,BEPC}$, $y_{p,CWEN}$, and $y_{p,NEE}$, sentiment scores related to inflation exhibit the highest uncertainty coefficient values. Lastly, when focusing on $y_{p,SHEL}$, sentiment scores for economy also possess among the highest uncertainty coefficients compared to other aspects.

In summary, sentiment scores concerning the economy and inflation consistently contain the most informative signals regarding various stock prices. Relative to sentiment scores for other aspects, they possess the greatest uncertainty coefficient values with respect to stock prices. Additionally, sentiment scores for other aspects, such as financial, china, stockmarket, interest, stock, tax, price, bitcoin, cost, bank, market, also exhibit elevated uncertainty coefficient values with stock prices. Notably, sentiment scores for financial show not only high uncertainty coefficient values but also a strong and explainable correlation and Granger causality with stock prices.