Analyzing Buffett Chairman Letters using Topic Models
I have had a fascination with Warren Buffett since I landed in the US and learned about him form a mentor of mine. Although I had heard of his famous chairman letters and how accessible they were, I was more interested in his bio and other content about him.
I finally broke down this year and started reading his letters which are publicly available on Berkshire Hathaway site. This holiday season, I went and tried to visualize the themes of his writing using Topic Models. What follows is a write up of the work I did.
Downloading data
I used wget and a simple python script to download the PDFs (since 1998) and extract the text using Ghostscript. For the plain HTML ones (pre1998) I used html2text and then extracted text. Since the writing were plain text and no fancy images etc. it was easy to get the data. Once extracted I had original letters in text file format from 1977-2014, a total of 38 files.
Learning about Topic Models
A separate technical note will follow this where I will document my learnings of this topic.
Topic Model tooling
I was back to using R, python to see what LDA/Topic Model library did the job I was looking for. I explored lda & topicmodel & tm libraries in R and LDA & gensim libraries in Python. In the end I ended up using Mallet from UMass which was simple and performant for the documents had. Both Python and R libraries took over 2 minutes to process the 38 docs for 500 iterations, while Mallet did the job in 10s of secs. A screenshot of using the tool on my Linuxtop.
Exploring dataset
Before I started looking at the topics and running LDA or one of its variants, I wanted to see how the word distribution looked like, so I used R to get the wordcloud which helped me to refine my stopword selection to be used in Mallet.
For completion, here the are the top 10 words in bar plot-
Topic based analysis of Buffet letters
Mallet is very easy to use and once you have the text documents primed, its a breeze. Since most of the cursory read of the letters suggested that there were similar themes that Buffett was writing, I wanted to understand if number of topics had any impact to how the themes evolved over the past 30+ years in Buffett’s writing. I tried 2,5,10,20,25,50 topics and ended up pruning 5,10,20 topics to see the evolution and did most of the analysis with 5 topics model. I show the HTML output produced by Mallet for 5 topics model with top 10 words
Then I proceeded to plot the topics over time for the 5,10,20 topics models. 20 topics model with top 3 topics per letter over time -
10 topics model with top 3 topics per letter-
5 topics model with top 3 topics per letter-
Moneyshot
My favourite tool from the viz toolchest has been multi-dimensional scaling (MDS) for higher dimensional data for the last decade and so I took the 5 topics model and tried to visualize the the letters over time using how each letter contributes to the topic (probablities). This produced the most cool graph clustering the 70s,80s,90s,2000s loosely.
