[2H6-OS-2c] OS-2

5:45 PM - 6:00 PM JST(8:45 AM - 9:00 AM UTC)

[2H6-OS-2c-02] Do Lazy Prices Hold in Japan?Vocabulary Direction in Corporate Disclosures and Stock Returns Using LLM-Assisted Dictionaries

Katsuhiko Okada¹, 〇Moe Nakasuji¹, Yasutomo Yamasaki¹, Takahiro Yamasaki² (1. Kwansei Gakuin University, 2. Osaka Sangyo University)

Keywords:

Text Mining,10K,LLM generated dictionary,Cosine Similarity,Jaccard Similarity

Cohen, Malloy, and Nguyen (2020) show that U.S. firms making larger 10-K text changes subsequently underperform, attributing this Lazy Prices effect to investor inattention. In the U.S., high litigation risk incentivizes defensive language, making text changes systematically negative. We test whether this holds in Japan, where shareholder litigation is rare and recoveries minimal. Using three text similarity measures on MD&A sections of annual securities reports for 12,191 firm-years across 1,436 firms (2015-2024), we find that overall disclosure changes do not predict returns. However, the direction of vocabulary change is informative. Using growth and risk dictionaries (40 terms each) constructed with LLM assistance, firms adding growth vocabulary significantly outperform, generating a value-weighted Carhart four-factor alpha of 5.77% per annum (t=2.61). This effect concentrates in large caps (6.05%, t=2.88) and is robust to translated Loughran-McDonald dictionaries (4.76%, t=2.29). Post-TSE reform subsamples show the effect broadening to equal-weighted portfolios (4.02%, t=3.34). The same textual signal carries opposite return implications depending on institutional context.