With a set goal in mind, we can narrow down potential data sources to investigate. Now the question becomes: is it necessary to collect the data of all fields available in a data source, or is a subset of attributes enough? It would be perfect if we knew in advance which attributes were key indicators or key predictive factors. However, it is in fact very difficult to ensure that the attributes hand-picked by a domain expert will yield the best prediction results. Hence, for each data source, it is recommended to collect all of the fields that are related to the project, especially in cases where recollecting the data is time consuming, or even impossible.

For example, in the stock price prediction example, we collected the data of all fields including open, high, low, and volume, even though we were initially not certain of how useful high and low predictions would be. Retrieving the stock data is quick and easy, however. In another example, if we ever want to collect data ourselves by scraping online articles for topic classification, we should store as much information as possible. Otherwise, if any piece of information is not collected but is later found valuable, such as hyperlinks in an article, the article might be already removed from the web page; if it still exists, rescraping those pages can be costly.

After collecting the datasets that we think are useful, we need to assure the data quality by inspecting its consistency and completeness. Consistency refers to how the distribution of data is changing over time. Completeness means how much data is present across fields and samples. They are explained in detail in the following two practices.