At each point information from neighbouring data is used. As the time series considered are finite, the areas at the start and end of the data (especially at higher period bands) will not have all the data required. One solution to make computation possible, chosen here, is to pad the time series with zeros where required. However, the zero padding will impact the reliability of the results. It is standard to use a cone of influence to represent this difference in reliability of results. Pale colours represent those areas outside the cone of influence with less reliable results (as seen on Fig 3). Higher period bands require more data for computation resulting in the cone shape.
While exact predictions are impossible, we asked five experts about what they’re paying attention to in the crypto space for the rest of 2021:

Posts per day and new authors can be retrieved from each subreddit programmatically; each post is timestamped, so historical time series can be generated by iterating through the posts. Posts per day are used here rather than comments per day. Each post on Reddit can have a number of associated comments in a one-to-many type relationship. However, examples exist where huge numbers of comments are generated that are unrelated to market activity; for example, sometimes people give away small amounts of cryptocurrency to everyone who comments with their public blockchain wallet address; this causes a huge spike in comments (wavelet coherence between comments per day and price were also generated, but as was expected showed less significant relationships than posts per day and price).
Every trade contributes towards auto-generating liquidity that goes into multiple pools used by exchanges