Research

This chapter examines whether news sentiment provides additional predictive information for cryptocurrency returns and volatility beyond that contained in market variables. I use asset-specific news headlines collected from GDELT to construct two daily sentiment measures based on a financial lexicon and FinBERT.The analysis covers Bitcoin, Ether, Solana, and XRP, with multipleforecast horizons. Seven forecasting models are considered, including linearregressions, tree-based methods, and a neural network.Forecasting performance is evaluated using an expanding-windowapproach, with RMSE comparisons and the Clark-West test.I consider both a common sample and asset-native samplesto examine the sensitivity of the results to the sampledesign. The analysis also examines whether the predictivecontribution of sentiment differs between high- andlow-volatility market conditions. The results indicate that news sentiment provides morefavorable improvements for volatility forecasting thanfor returns. Statistically supported return gains arerelatively limited and rarely persist across the twosample designs. In contrast, several volatility forecastsbenefit from sentiment information, with the largestimprovements generally obtained using Multilayer Perceptron(MLP), particularly for XRP. However, some of these gainsweaken or reverse when different sample periods areconsidered. The results also show that improvements occurunder different market volatility conditions, without a consistent advantage for either regime.