• Take ~1800 series on NASDAQ North US between 2008 and 2015
• Run some "magic" feature engineering (c.f. @minhtran) to create 28 features for each symbol
• Split all the stock into sequences of length 40 (days)
• Train a Seq2Seq model to predict the stock feature vector of the next day, given the input.
• The final hidden state (of the encoder) is used as a "vector representation" of the 40-day segment. The dimension of this vector is equal to the number of hidden units in the RNN (1024 in our case)
• t-SNE all the hidden states.

To train the RNN:

$ python train.py --learning_rate=0.1

After training, the feature vectors can be extracted by:

$ python train.py --decode

We trained a Seq2Seq model of 1024 GRU units, with learning rate of 0.1. The final visualization can be seen here

• Use LSTM instead of GRU.
• Use a more sophisticated Seq2Seq model (i.e. more layers, attention models...)
• Maybe some clever strategy in splitting the series/feature engineering
• Fix bugs (if any)