Quicktake: BPE, WordPiece, and SentencePiece
One table to compare popular tokenization methods: BPE, WordPiece, and SentencePiece.
Yu Zhu "祝宇"
Assistant Professor, Information Systems
University of Delaware, USA
Ph.D., Information Systems
University of Utah, USA
Ph.D., Finance
Zhejiang University, China
Implementation details: From the original Transformer to GPT
This article comapres the implementation details between the original Transformer and GPT. These tricks are critical to performance but not always explained in the paper.
Transformers (deep learning models) are better at predicting the tail distribution
In a setting of PEAD (post-earnings-announcement-drift) prediction using earnings call transcripts, I found Transformers (deep learning models) have a larger performance lead on extreme data points (data at the tails of the distribution).
(R) Super Efficient Event Study Code
A highly efficient (>10x speedup) and concise (core part is less than 40 lines) event study R program that replicates WRDS’s SAS version.
All you need to know about data/time in R and SAS
How date and time are stored/processed in R and SAS, and how to use them wisely in WRDS or in your daily programming.
Git tips
As a Git non-prouser, I found these tips very helpful.
Some facts on Earnings Announcement and Analyst Revision
Some facts on earnings announcement and analyst revision that you may not know before.
Expected Return Model
How to choose the right expected return model?