THE TODO LIST:
-
(DONE) Factor standardization use scikit learn
-
(DONE) Change Y values to shifted X values for sequence learning
-
(DONE) Does MSE need to be divided by batch size batch? No. Done in TF.
-
(DONE) Implement Clairvioant and Naive models
-
(DONE) Implement batch sequences that only require final step being an active stock.
-
(DONE) Implement predicting next n-timestep average of inputs in batch_generator
-
(DONE) Incorporate merge-model-with-simdata.pl into euclid2
-
(DONE) Simulate clairvoyant progression from 0,3,6,12, ... months to be how perf improves
-
(DONE) In predict.py, make predictions even when there is no target data available
-
(DONE) Create file cache (pickle) for batch_generator
-
(DONE) Add auxilary input features -- ones that are not predicted/targets (e.g., momentum)
-
(DONE) Re-working scaling/unscaling implementation so it is more intuitive
-
(DONE) Layer normalization in RNN
-
(DONE) In predict.py, output predictions timesteps less than t. I.e., t-1, t-2, 0
-
(DONE) Implement variable length sequences
-
(DONE) RNN cost function upweight last k time steps instead of just last time step
-
Configurable validation/holdout set methodology (holdout time window or companies)
-
Implement a genetic aglorithm for hyper-parameter space search
-
max-norm regularization for RNN and MLP (use tf.clip_by_norm)
-
Trainable ReLu units in MLP
-
rename config's nn_type to model_type
-
Documentation. Starting with README.md
-
Make caching faster