Class

https://courses.nvidia.com/dashboard

Jupyer notebooks: Shift+Enter to run cell

Linguistic Concepts

• conference - anaphors
• gang of four design
• null subject
• recursion

Word Embeddings

• HMMS, CRF, PGMs
  • CBoW -Bag of Words / ngrams - feature per word/n items
  • 1-hot Sparse input - create a vector the size of the entire vocabulary
• Stop Words
• TF-IDF

Word2Vec

Skip-Gram

• Firth 1957 Distributional Hypothess

• Word Cloud

Financial News

Yuval Mazor NVIDIA https://www.google.com/search?q=Yuval+Mazor+%40nividia.com&oq=Yuval+Mazor+%40nividia.com&aqs=chrome..69i57j69i58.9126j0j8&sourceid=chrome&ie=UTF-8

Tools:

• Glove
  • dot product
• FastText
  • Skipgram
  • Continuous bag of words

Multi-channel LSTM Network Keras wih TensorFlow Utilize the GloVe and FastText Skipgram pretrained embeddings, allows he underlying network to access larger feature space to build complex features on top of.

Can use utilize combinations of various corpus and embedding methods for better performance

Bidirectional LSTM network is used o encode sequential information on the embedding layers.

Dense layer to project fnal output classification

Use embedding... embeddings = transfer learning

? CNN vs BI-LSTM (RNN) this approach, BI-LSTM does not need a lot of data

Attention mechanism -- translate ... you can look back

                   ... not a fixed vector size

• GloVe
• Fasttext
• News articles per day
• News data source
• Word embeddings
• Natural Language Processing
• Sentiment Analysis

Deep Autoencoders for Anomaly Detection

Variable Auencoders

• clustering - latent layers may tell you what number of clusters
• anomaly detection

https://courses.nvidia.com/courses/course-v1:DLI+L-FI-06+V1/info

PCA or TSenee

Statistical Arbitrage

arbitrage - monies, stocks (price is better than it should be - fair market value) how right, or how rich? Mean inversion Autoencoder learn the fair market value, then feed in current value

reconstruction error is a signal - just one signal, consider a basket of signals

backtesting - if we run this on previous historical events how well does our algorithm work? (don't use training data !!)