JS
J.R. Sijsenaar
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
A single social media post can trigger a 10% drop in an asset’s price within a day, as demonstrated by Elon Musk’s ”Tesla stock price is too high imo” post. With the growth of social media and advances in language models, sentiment analysis becomes increasingly relevant and computationally achievable in financial markets. From a behavioral finance perspective, it is a well-established result that asset price changes stemming from sentiment shifts exhibits meanreversion. This property motivates the construction of a risk measure by separating sentimentinduced excess volatility from the total volatility.
This thesis aims to quantify this sentiment-induced volatility using a dataset of StockTwits posts from Kaggle. We began by analysing the data, where it was revealed that the posting activity displays properties that are associated with information diffusion, including heavy-tailed activity, burstiness and temporal autocorrelation. Accordingly, FinBERT and TwitterRoBERTa were employed to classify the posts as bullish, bearish or neutral, and the classifications were compared to the user-provided labels. These two BERT-based language models provide a wide coverage by specializing in Finance and colloquial language, respectively.
The predictive power of the resulting sentiment features was evaluated using (Bi-)LSTM and (Bi-)GRU. This power was measured by comparing only lagged-returns models to models enriched with attention-, calendar- and sentiment-based features. It was found that the returns-only models outperformed the enriched feature sets across all five assets and two distinct evaluation approaches. Inspecting the learning and validation curves further showed that the enriched features reduced model generalizability, rather than improving predictive performance.
The findings do not indicate that the constructed sentiment features provide any additional predictive information within the proposed framework. As a result, the evidence does not support proceeding with the construction of the proposed risk measure. This could potentially be the result of the choice of the selected target variable, log-squared returns, which is closely related to the lagged returns. Consequently, future research could consider a different volatility proxy; one that considers intraday price ranges. Such a quantity would not be as directly related to lagged returns and may allow sentiment-based features to contribute more meaningfully to the modeling process. ...
This thesis aims to quantify this sentiment-induced volatility using a dataset of StockTwits posts from Kaggle. We began by analysing the data, where it was revealed that the posting activity displays properties that are associated with information diffusion, including heavy-tailed activity, burstiness and temporal autocorrelation. Accordingly, FinBERT and TwitterRoBERTa were employed to classify the posts as bullish, bearish or neutral, and the classifications were compared to the user-provided labels. These two BERT-based language models provide a wide coverage by specializing in Finance and colloquial language, respectively.
The predictive power of the resulting sentiment features was evaluated using (Bi-)LSTM and (Bi-)GRU. This power was measured by comparing only lagged-returns models to models enriched with attention-, calendar- and sentiment-based features. It was found that the returns-only models outperformed the enriched feature sets across all five assets and two distinct evaluation approaches. Inspecting the learning and validation curves further showed that the enriched features reduced model generalizability, rather than improving predictive performance.
The findings do not indicate that the constructed sentiment features provide any additional predictive information within the proposed framework. As a result, the evidence does not support proceeding with the construction of the proposed risk measure. This could potentially be the result of the choice of the selected target variable, log-squared returns, which is closely related to the lagged returns. Consequently, future research could consider a different volatility proxy; one that considers intraday price ranges. Such a quantity would not be as directly related to lagged returns and may allow sentiment-based features to contribute more meaningfully to the modeling process. ...
A single social media post can trigger a 10% drop in an asset’s price within a day, as demonstrated by Elon Musk’s ”Tesla stock price is too high imo” post. With the growth of social media and advances in language models, sentiment analysis becomes increasingly relevant and computationally achievable in financial markets. From a behavioral finance perspective, it is a well-established result that asset price changes stemming from sentiment shifts exhibits meanreversion. This property motivates the construction of a risk measure by separating sentimentinduced excess volatility from the total volatility.
This thesis aims to quantify this sentiment-induced volatility using a dataset of StockTwits posts from Kaggle. We began by analysing the data, where it was revealed that the posting activity displays properties that are associated with information diffusion, including heavy-tailed activity, burstiness and temporal autocorrelation. Accordingly, FinBERT and TwitterRoBERTa were employed to classify the posts as bullish, bearish or neutral, and the classifications were compared to the user-provided labels. These two BERT-based language models provide a wide coverage by specializing in Finance and colloquial language, respectively.
The predictive power of the resulting sentiment features was evaluated using (Bi-)LSTM and (Bi-)GRU. This power was measured by comparing only lagged-returns models to models enriched with attention-, calendar- and sentiment-based features. It was found that the returns-only models outperformed the enriched feature sets across all five assets and two distinct evaluation approaches. Inspecting the learning and validation curves further showed that the enriched features reduced model generalizability, rather than improving predictive performance.
The findings do not indicate that the constructed sentiment features provide any additional predictive information within the proposed framework. As a result, the evidence does not support proceeding with the construction of the proposed risk measure. This could potentially be the result of the choice of the selected target variable, log-squared returns, which is closely related to the lagged returns. Consequently, future research could consider a different volatility proxy; one that considers intraday price ranges. Such a quantity would not be as directly related to lagged returns and may allow sentiment-based features to contribute more meaningfully to the modeling process.
This thesis aims to quantify this sentiment-induced volatility using a dataset of StockTwits posts from Kaggle. We began by analysing the data, where it was revealed that the posting activity displays properties that are associated with information diffusion, including heavy-tailed activity, burstiness and temporal autocorrelation. Accordingly, FinBERT and TwitterRoBERTa were employed to classify the posts as bullish, bearish or neutral, and the classifications were compared to the user-provided labels. These two BERT-based language models provide a wide coverage by specializing in Finance and colloquial language, respectively.
The predictive power of the resulting sentiment features was evaluated using (Bi-)LSTM and (Bi-)GRU. This power was measured by comparing only lagged-returns models to models enriched with attention-, calendar- and sentiment-based features. It was found that the returns-only models outperformed the enriched feature sets across all five assets and two distinct evaluation approaches. Inspecting the learning and validation curves further showed that the enriched features reduced model generalizability, rather than improving predictive performance.
The findings do not indicate that the constructed sentiment features provide any additional predictive information within the proposed framework. As a result, the evidence does not support proceeding with the construction of the proposed risk measure. This could potentially be the result of the choice of the selected target variable, log-squared returns, which is closely related to the lagged returns. Consequently, future research could consider a different volatility proxy; one that considers intraday price ranges. Such a quantity would not be as directly related to lagged returns and may allow sentiment-based features to contribute more meaningfully to the modeling process.
We inspect the behavior of the probability that a weighted sum of random variables with log-normal tails is greater than its expected value. Under the right conditions for the weights and the variance being set to 1; we were able to bound a suitable transformation of this probability with the upper bound being a fixed factor of the square root of e above the lower bound. Beyond this, we analyse the conditions on the weights and determine a method for letting the weights be random and give an example.
We end off by extending our result to general variance, where we see that the deviation between the lower and upper bound as well as the domain for the result are dependant on the variance. ...
We end off by extending our result to general variance, where we see that the deviation between the lower and upper bound as well as the domain for the result are dependant on the variance. ...
We inspect the behavior of the probability that a weighted sum of random variables with log-normal tails is greater than its expected value. Under the right conditions for the weights and the variance being set to 1; we were able to bound a suitable transformation of this probability with the upper bound being a fixed factor of the square root of e above the lower bound. Beyond this, we analyse the conditions on the weights and determine a method for letting the weights be random and give an example.
We end off by extending our result to general variance, where we see that the deviation between the lower and upper bound as well as the domain for the result are dependant on the variance.
We end off by extending our result to general variance, where we see that the deviation between the lower and upper bound as well as the domain for the result are dependant on the variance.