This is a toy project by a human learning to do some practical machine-learning. The model used is just a simple LSTM and a Dense layer.
The initial implementation splits the data by stock_id into training and validation batches. The losses decreased nicely, but the resulting model is not good at predicting real world data.
The current implementation splits the data by time, reserving the first N timesteps for training, and the rest for validation. The result is that the training losses decrease nicely, and the validation losses stayed the same.
In other words... a nice exercise but it DOES NOT WORK at predicting real world data.
Input data should be in the format of:
stock_id,days_1,begin_price,end_price,low_price,high_price,b_large_volume,b_big_volume,b_middle_volume,b_small_volume,s_large_volume,s_big_volume,s_middle_volume,s_small_volume
000800.SZ,2010-09-07 00:00:00,15.18,14.87,14.82,15.33,21570897,1445674,844167,0,21797815,996369,1066349,205
000800.SZ,2010-09-08 00:00:00,15.00,14.88,14.67,15.04,15448956,925000,1049072,0,16144750,617026,661073,179
000800.SZ,2010-09-09 00:00:00,14.65,14.42,14.27,14.80,20986988,1377623,1413536,0,21996045,840644,941249,209
000800.SZ,2010-09-10 00:00:00,14.55,14.08,14.00,14.62,15330849,1251200,871800,0,15859016,687931,906769,133
prepare-data.py is used to convert data into usable batches.
run-keras.py and run-torch.py are used to fit the data.
For all commands use --help to see parameters. Default parameters can also be changed by changing constants at the top of the files.
Prepare data into train/val batches:
python prepare-data.py -f data/input.csv
Training (w/ validation) from scratch
python run-keras.py
Training from previous checkpoint
python run-keras.py -l 300
where 300 is the checkpoint saved after number of episodes of training.
Validation only
python run-keras.py -m v -l 300
Prepare data into pred batches:
python prepare-data.py -f data/input.csv -yl 0
Prediction
python run-keras.py -l 300 -m p
where 300 is the checkpoint saved after number of episodes of training.