Classification Forex Prediction Model

classification-forex-model.py

import csv
import random

import numpy as np
import matplotlib.pyplot as plt

import math

import os
os.environ["THEANO_FLAGS"] = "mode=FAST_RUN,device=gpu,floatX=float32,allow_gc=False"

from keras.models import Sequential
from keras.layers import Dense, LSTM, Flatten, TimeDistributed, Reshape
from keras.callbacks import ReduceLROnPlateau, ModelCheckpoint
from keras.layers.convolutional import Convolution2D

LINES_OF_DATA = 100000

PERIODS_PER_X = 180
PERIODS_FORWARD = 60

PIP_LEVEL = 10

NEG_PIP = [1, 0, 0]
NIL_PIP = [0, 1, 0]
POS_PIP = [0, 0, 1]

FILE_NAME = 'with_y_simple_all.csv'

# set up lists to contain data
raw_data = []

X = []
Y = []

print 'loading data'

f = open(FILE_NAME)
for line in f:
	raw_data.append([float(x) for x in line.split(',')])
	if len(raw_data) % 10000 == 0 and len(raw_data) > 0:
		print 'loaded ' + str(len(raw_data))
	if len(raw_data) >= LINES_OF_DATA:
		break
f.close()

FEATURES = len(raw_data[0])
print raw_data[0]

# break raw_data into working array 
print 'processing and shaping data'

for x in xrange(len(raw_data)-PERIODS_PER_X-PERIODS_FORWARD-1):
	X.append(raw_data[x:x+PERIODS_PER_X])


for x in xrange(len(X)):
	classifier_value = NIL_PIP
	for y in xrange(PERIODS_FORWARD):
		if X[x][-1][3] - raw_data[PERIODS_PER_X+x+y-1][3] >= PIP_LEVEL*0.0001:
			classifier_value = POS_PIP
			break
		elif X[x][-1][3] - raw_data[PERIODS_PER_X+x+y-1][3] <= -PIP_LEVEL*0.0001:
			classifier_value = NEG_PIP
			break
	Y.append(classifier_value)

print np.array(X).shape
print FEATURES

print 'setting up model'

model = Sequential()
# model.add(Reshape((1, PERIODS_PER_X, FEATURES), input_shape=(PERIODS_PER_X, FEATURES)))
# model.add(Convolution2D(64, 3, 3, border_mode='same'))
# model.add(Convolution2D(1, 1, 1, border_mode='same'))
# model.add(Reshape((PERIODS_PER_X, FEATURES)))
model.add(LSTM(PERIODS_PER_X, input_dim=FEATURES, input_length=PERIODS_PER_X, return_sequences=True))
model.add(LSTM(PERIODS_PER_X, input_dim=FEATURES, input_length=PERIODS_PER_X, return_sequences=True))
model.add(TimeDistributed(Dense(FEATURES)))
model.add(Flatten())
# if doesn't fit as well as other one, try adding more layers, or try using the old model with only price information (KISS)
model.add(Dense(3, activation='softmax'))

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

filename = 'classification-model.h5'
save_best_model = ModelCheckpoint(filename, monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto')

model.fit(X, Y, batch_size=100, nb_epoch=2500, verbose=1, callbacks=[save_best_model], validation_split=.3, validation_data=None, shuffle=True, class_weight=None, sample_weight=None)

print 'saving model'
model.save(filename)
print 'model saved'