4 ms·
Disruptive Innovation and IPO Outcomes: Evidence from Machine Learning
- Bostonian 3y agoquoting the paper: We use keywords for 29 disruptive technologies identified by Bloom et al. (2021) and a semi-supervised machine learning method proposed by Li et al. (2021a) to quantitatively measure the discussion of disruptive innovation in 3,440 IPO prospectuses between 1994 and 2021. These advanced technologies have dominated discussions between firm executives and investors in earning calls over the last twenty years (see Bloom et al. 2021), so it is reasonable to expect them to be discussed in the prospectus. We train a machine learning model, word2vec, on the text of IPO prospectuses and use Bloom et al.’s 29 technology keywords as seed words. We compute the cosine similarities between the seed word vector and each unique word in IPO prospectuses. We use the 500 words that have the highest cosine similarity scores with the seed words to construct our context-based dictionary of disruptive innovation. Our expanded dictionary, shown in Appendix 3, identifies additional technologies beyond the 29 key technologies identified by Bloom et al. (2021), e.g., 128bit, airbag, analog converter, and laser. We then compute a disruptive innovation score (DIS) for an IPO firm as the weighted count of the words from our disruptive innovation dictionary that appear in the IPO prospectus divided by the total number of words in the prospectus, expressed as a percentage. The weights of words in the numerator of DIS are computed using the WF.IDF method, which gives lower weight to words that appear more frequently in the prospectus, as described in section 3. DIS captures firms’ prior and future adoptions of the disruptive technologies discussed in IPO prospectuses.