Tribots read and write
10-Ks, annual financial corporate reports, is the basis of intelligent investment. These reports often reveals key information about the company.
Intro to Word to Vec
Word2Vec is a method that would try to represent the document into a vector form
General idea is that for the vocabularies used in the document, one would come up with a probability distribution of words that would be used in a nearby expression
2 Methods for doing this is continuous-bag-of-words and skip-gram
Continuous-bag-of-words tries to produce probability distribution of a word given a list of words
Skip-gram tries to provide probability distribution of words that show up in similar content given a vocabulary
Authorship attribution
We have also tried to apply authorship attribution analysis to the reports
It seems that doc2vec analysis is not quite effective at distinguishing authors unless the writings are drastically different
One methodology is to use key features that would give characteristics about the authors and use k-means clustering
Lexical Features:
average number of words per sentence
sentence length variation(standard deviation of words per sentence)
Lexical diversity(number of unique words/words used in document)
Syntactical:
Frequencies for common Parts of Speech types(singular/plural noun, proper noun, determiner, preposition/conjunction,adjective)
Bag of Words:
Count the most common words in the documents and apply clustering
| Misclassification Rate | A Tale of Two Cities vs. The Great Gatsby | The Great Gatsby vs
Sherlock Holmes |
A Tale of Two Cities vs. Shakespeare | Amazon 10K vs.
Altria 10K |
| Lexical | 14% | 0% | 46% | 0% |
| Syntax | 14% | 25% | 46% | 0% |
| BOW | 46% | 38% | 46% | 0% |
Good and Bad Companies Classification
If the company’s stock price has decreased by 50% we have considered them “bad”
If the company’s stock price has increased by more than 100% we have considered them “good” (over 5 years)
Random Forest For Doc2Vec
| Confusion Matrix | Prediction
Losers |
Prediction
Winners |
| Actual
Losers |
11 | 33 |
| Actual
Winners |
5 | 130 |
Logistic Classification For Doc2Vec
| Confusion Matrix | Prediction
Losers |
Prediction
Winners |
| Actual
Losers |
24 | 20 |
| Actual
Winners |
36 | 109 |
Recurrent Neural Networks
Neural Network has been around since 1940s but were not so useful until about 5 years ago
Whereas regular Neural Networks does not have a sense of time, Recurrent Neural Networks try to capture time element while working with neural networks by using previous output as another input
Recurrent Neural Networks by itself is not as effective so people have incorporated LSTM to forget or remember previous outputs
Generating Corporate Report using RNN/LSTM
10K reports were too large for us to train using RNN
Following the example of writing a similar story in Aesop’s fable, we have written summary using company profiles in CNBC
“the bell outwit met . nobody will all mouse got up and said that is all very well , but he thought would meet the case . you will all agree , said attached chief'”(Aesop’s Fable LSTM RNN output)
“other service offerings for play and retailers through the consumer and sale beverages. Hasbro, The Investment segment focuses sales and facilitates that sectors. Wholesale in food countries in China its in fundamental storage, equity, and”(somewhat makes sense)
Conclusion
While we made some progress in using natural language processing to investigate 10-Ks, here are some future research direction:
1. Further analysis is needed to examine the relationship between long-term company health with different sections of 10-K (i.e., risks, business overview)
2. Features generated by NLP alone might not have the necessary predicative power and therefore more quantitative measures, such as valuation models should be included as well
3. 10-Qs, quarterly financial reports, are filed after each quarter and might better reflect company’s current status.
The future of using NLP to understand financial reports are unbound. This similar analysis can be applied to other financial documents that might contain crucial information for intelligent investment. By combining NLP with recurrent neural network (RNN) and long and short term memory (LSTM), machine learning might be able to automatically compose a 10-K!
About Authors
Daniel Rim
View all posts by Daniel Rim >
Jason Chiu
View all posts by Jason Chiu >