By Christoffer and Sebastian

Repo mainly made for sharing, not direct implementation of code base as plenty of refractoring needed to archive that standard.

Code and papers writting for courses Empirical Asset Pricing and Economic Forecasting. Both papers utilize sentiment extracted from Reddit comments collected from google big query.

Requires setting up own google cloud account to access big query and downloading the json credential file. See Kode.Functions.collect_big_query

This paper examines the cryptocurrency market with an outset in traditional empirical asset pricing, to assess potential similarities with the asset classes for which our theory is based, namely equities. More specifically, we investigate the potential role sentiment plays in this market by analyzing the significance of the excess returns on a zero-investment strategy based on daily sentiment sorted portfolios. We then examine if this cross-sectional cryptocurrency return predictor can be explained using a low number of common factors, utilizing the Fama-Macbeth regression setup and related factors [Fama & MacBeth, 1973], [Fama & French, 1992]. Finally, we construct a sentiment factor and test its relevance in risk pricing when controlling for established factors. Portfolios and common factors are constructed using data from 01.01.14-29.30.19, consisting of twelve cryptocurrencies which historically have constituted at least 80% of the entire market capitalization. Computational restrictions have resulted in the use of a relatively low amount of coins, which entails that the constructed portfolios likely suffer from a bit of noise and idiosyncratic risk. We have opted to use quartile portfolios to balance this noise while still obtaining a proper amount of cross-sectional variability in portfolio returns.

This paper seeks to determine the performance of the most prevalent forecasting models when incorporating social media sentiment within the field of Bitcoin. We further conduct forecast combinations, in order to examine whether optimal performance could be obtained by utilizing both the econometric models and the machine learning techniques with and without sentiment.