The classification of affect as positive (as in "good") or negative (as in "bad") is called valence.
There is a large body of research regarding word valence, owing to interest from corporate communications and public messages, for example as delivered through social media. To further address demand, Bradley & Lang (1999) developed a database of dictionary words and their valence scores, that is available for download. More recently, such databases have been developed for dozens of languages, and the English database has been greatly expanded as well. These databases often feature multiple dimensions of affect, but if you are not interested in them, then just look at valence and ignore the rest.
With such databases in hand, it is possible to test how different factors predict word valence. Many hypotheses have been tested, such as word length, conjugations, age of acquisition, etc. A comprehensive review by Warriner et al (2013) lists a few more examples of correlations with lexical properties, such as smell, color, and motion:
Most correlations that emotional ratings show with other semantic
properties are weak to moderate, with the exception of correlations
with variables that directly tap into emotional states.
Of particular note, and extensively studied in this field, are concreteness, imageability, context availability, and familiarity. Familiarity for example, is a term that has to do with how well known and how common words are - common words tend to be rated more positively than uncommon ones. This has to do with the mere-exposure effect, and more generally, processing fluency (a.k.a. "cognitive ease"):
Fluency and familiarity have been shown to lead to the mere exposure
effect. Research has found that repetition of a stimulus can lead to
fluent processing which leads to a feeling of liking. ... Later
research observed that high perceptual fluency increases the
experience of positive affect.
Another good list of databases and research on concreteness, imageability, context availability, and familiarity can be found in Riegel et al (2015). As I said, it's a big field, but hopefully this gets you started.