[Paper Review] Beyond Trading Data: The Hidden Influence of Public Awareness and Interest on Cryptocurrency Volatility

Zeyd Boukhers, Azeddine Bouabdallah+2 more|arXiv (Cornell University)|Feb 12, 2022

Stock Market Forecasting Methods4 citations

TL;DR

This paper proposes an AdaBoost-LSTM ensemble model that integrates trading data, social media sentiment, blockchain metrics, and search volume to forecast Bitcoin prices and their volatility distribution. The multimodal approach achieves a 19.29% improvement over baseline methods, with a test MAE of $132.03, demonstrating superior accuracy and enhanced decision support through uncertainty-aware predictions.

ABSTRACT

Since Bitcoin first appeared on the scene in 2009, cryptocurrencies have become a worldwide phenomenon as important decentralized financial assets. Their decentralized nature, however, leads to notable volatility against traditional fiat currencies, making the task of accurately forecasting the crypto-fiat exchange rate complex. This study examines the various independent factors that affect the volatility of the Bitcoin-Dollar exchange rate. To this end, we propose CoMForE, a multimodal AdaBoost-LSTM ensemble model, which not only utilizes historical trading data but also incorporates public sentiments from related tweets, public interest demonstrated by search volumes, and blockchain hash-rate data. Our developed model goes a step further by predicting fluctuations in the overall cryptocurrency value distribution, thus increasing its value for investment decision-making. We have subjected this method to extensive testing via comprehensive experiments, thereby validating the importance of multimodal combination over exclusive reliance on trading data. Further experiments show that our method significantly surpasses existing forecasting tools and methodologies, demonstrating a 19.29% improvement. This result underscores the influence of external independent factors on cryptocurrency volatility.

Motivation & Objective

• To identify which data modalities—trading data, social media sentiment, blockchain metrics, and search volumes—most directly influence cryptocurrency prices.
• To evaluate whether ensemble learning on multimodal data enhances forecasting performance compared to unimodal or single-model approaches.
• To develop a forecasting system that provides not only predicted prices but also fluctuation distributions to convey prediction confidence.
• To support investor decision-making with reliable, interpretable, and uncertainty-aware short-term price forecasts.

Proposed method

• The model employs an AdaBoost-LSTM ensemble architecture that combines multiple weak LSTM learners trained on different data modalities.
• Input features include historical price data, social media sentiment scores, blockchain metrics (e.g., hash rate, network difficulty), and Google search volumes.
• Each modality is preprocessed and normalized before being fed into separate LSTM networks, which are then ensembled using AdaBoost to improve generalization.
• The final prediction integrates outputs from all modalities, with uncertainty modeled via predicted fluctuation distributions for each forecasted day.
• The model is trained and evaluated on a dataset spanning January 2020 to July 2020, with performance measured using MSE, RMSE, and MAE.
• For long-term forecasting, predictions are recursively generated by appending each predicted value to the input sequence, simulating multi-day forecasts.

Experimental results

Research questions

• RQ1 RQ1: Which combination of data modalities—trading data, social media sentiment, blockchain information, and search volumes—has a direct and measurable impact on cryptocurrency price movements?
• RQ2 RQ2: Is ensemble learning on multimodal data effective in improving the accuracy and robustness of cryptocurrency price forecasting compared to unimodal or single-model approaches?
• RQ3 RQ3: How can price forecasting be enhanced by providing not only a point estimate but also a distribution of potential price fluctuations to convey model confidence?

Key findings

• The proposed AdaBoost-LSTM ensemble model achieved a test Mean Absolute Error (MAE) of $132.027, representing a 19.29% improvement over the best baseline method.
• The model’s MAE distribution showed that 68% of predictions were within ±$500 of the actual Bitcoin price, indicating strong short-term forecasting accuracy.
• The inclusion of multimodal data—especially social media sentiment and search volumes—significantly improved forecasting performance compared to models relying solely on trading data.
• The model’s predicted fluctuation distributions effectively conveyed prediction confidence, with flatter distributions indicating high uncertainty during volatile market conditions.
• Long-term forecasts showed a performance degradation over time, with the average MAE increasing from $132.03 on day 1 to higher values by day 30, as expected due to error accumulation.
• An open-source implementation of the model is provided to support reproducibility and future research.