I am trying to implement an evolving neural network on time series Forex data where the model will receive as inputs 3 different exchange rates on a particular timeframe and the base currency will be the same in all 3 inputs (e.g. USD/CHF, USD/JPY and USDZAR all have the same base currency namely USD when used as inputs). The network would then have to predict the direction of the base currency of the next weeks or days (based on timeframe used in data) using based on the inputs. It has the following options for prediction to choose from:
- Sell
- Strong Sell
- Buy
- Strong Buy
I have tried implementing this in the NEAT python framework based on the simplest example at https://neat-python.readthedocs.io/en/latest/xor_example.html using a very similar config file, representing "Sell" as "0", "Strong Sell" as "1", "Buy" as "2" and "Strong Buy" as "3" in the code. However I haven't been successful since the maximum fitness level I could get on the training set of 84 observations about 75%, this is after I decide to change the NEAT to predict only "Sell" and "Buy" represented by "0" and "1" respectively. I have tried implementing this using a recurrent as well as feedforward type network, not much improvement took place. I have also tried modifying the mutation rates. Kindly assist me with increasing the models fitness level as well as prediction accuracy on the test set. See code below and Thank you.
The data used below is data used here runs over 3 years.
from __future__ import print_function
import neat
import visualize
import random
# 3-input, inputs and expected outputs(training set).
xor_inputs = [(1.1241,1.2976,1.02923013585838),
(1.1201,1.2434,1.0231225700839),(1.0971,1.2187,1.00999899000101),
(1.0884,1.2235,1.00644122383253),(1.0987,1.2185,1.01142914938809),
(1.1138,1.2519,1.03241792277514),(1.0856,1.2602,1.01163378856854),
(1.0588,1.2346,0.99009900990099),(1.0587,1.2476,0.986679822397632),
(1.0673,1.2729,0.988630746416214),(1.0562,1.2576,0.982318271119843),
(1.0451,1.2497,0.974563882662508),(1.0456,1.2292,0.973994350832765),
(1.0516,1.2338,0.981932443047918),(1.0533,1.2287,0.982028871648826),
(1.0644,1.2179,0.990982063224656),(1.0703,1.2375,0.998203234178479),
(1.0699,1.2555,1.000100010001),(1.0784,1.2488,1.00745516824501),
(1.0641,1.2491,0.997506234413965),(1.0614,1.2414,0.997207818109294),
(1.0563,1.2468,0.992457324335054),(1.0623,1.2295,0.992457324335053),
(1.0672,1.2171,0.989119683481701),(1.0738,1.2394,1.00130169219986),
(1.0799,1.247,1.00867460157353),(1.0652,1.255,0.996909580301067),
(1.059,1.2373,0.991080277502478),(1.0612,1.2524,0.994629003381739),
(1.0728,1.2813,1.00411687920474),(1.0897,1.2951,1.0048231511254),
(1.0998,1.2981,1.01317122593718),(1.0931,1.2891,0.99930048965724),
(1.1207,1.3035,1.02838338132456),(1.1183,1.2804,1.02616726526424),
(1.1282,1.2887,1.0391769718383),(1.1196,1.2745,1.03135313531353),
(1.1198,1.2776,1.02690490860546),(1.1194,1.2718,1.03145951521403),
(1.1426,1.3027,1.04351455702807),(1.1401,1.2892,1.03734439834025),
(1.1469,1.3096,1.03788271925272),(1.1663,1.2995,1.05741778576716),
(1.1752,1.3135,1.03220478943022),(1.1772,1.3036,1.02785486689279),
(1.1821,1.3012,1.03960910697578),(1.1762,1.2876,1.03626943005181),
(1.1926,1.2887,1.04515050167224),(1.1859,1.2951,1.03659168653467),
(1.2035,1.3198,1.05876124933827),(1.1943,1.3593,1.0417751849151),
(1.1954,1.3493,1.03177878662815),(1.1814,1.3397,1.03284445362528),
(1.1734,1.3066,1.02249488752556),(1.1823,1.3287,1.02616726526424),
(1.1785,1.319,1.01595042161943),(1.161,1.3129,1.00230530219505),
(1.1609,1.3076,0.99930048965724),(1.1665,1.3191,1.00421771440048),
(1.1795,1.3214,1.01153145862836),(1.193,1.3337,1.02061645233721),
(1.1891,1.3473,1.02396067990989),(1.1764,1.3392,1.00704934541793),
(1.1754,1.3321,1.00959111559818),(1.1859,1.3362,1.01081572829273),
(1.1998,1.3515,1.02627257799672),(1.203,1.3571,1.02543068088597),
(1.2187,1.3729,1.03284445362528),(1.2222,1.3852,1.03842159916926),
(1.2421,1.417,1.07100781835707),(1.2462,1.4123,1.07376785139053),
(1.2235,1.3838,1.06371662589086),(1.2406,1.4042,1.07781849536538),
(1.2293,1.3967,1.06746370623399),(1.2317,1.3805,1.06587081645705),
(1.2307,1.3849,1.05130361648444),(1.2289,1.3941,1.0501995379122),
(1.2353,1.4134,1.05574324324324),(1.2323,1.4018,1.0481081647626),
(1.2283,1.4092,1.04242676951944),(1.2331,1.4241,1.03906899418121),
(1.2288,1.4004,1.02606197414324),(1.213,1.378,1.01224820325944),
(1.196,1.3533,1)]
xor_outputs = [(3,),(3,),(2,),(0,),(1,),(2,),(3,),(2,),(1,),(3,),(3,),
(2,),(1,),(0,),(0,),(1,),(0,),(0,),(2,),(3,),(2,),(2,),(2,),(1,),(1,),
(2,),(3,),(1,),(1,),(1,),(1,),(3,),(1,),(3,),(1,),(3,),(0,),(2,),(1,),
(3,),(1,),(0,),(0,),(2,),(0,),(3,),(1,),(2,),(1,),(2,),(2,),(2,),(3,),
(1,),(3,),(3,),(3,),(1,),(1,),(1,),(0,),(3,),(2,),(1,),(1,),(0,),(1,),
(1,),(1,),(0,),(3,),(1,),(3,),(2,),(2,),(2,),(1,),(3,),(2,),(0,),(3,),
(3,),(3,),(2,)
]
# 3-input, inputs and expected outputs(test set).
xor_inputs2 = [(1.1944,1.3543,0.99940035978413),
(1.1778,1.3474,1.00220485067148),(1.1652,1.3309,1.01030511214387),
(1.1661,1.3348,1.0120433154539),(1.1768,1.3412,1.01502233049127),
(1.1609,1.3285,1.00240577385726),(1.1657,1.327,1.01214574898785),
(1.1685,1.3209,1.00938730190774),(1.1747,1.3285,1.01020305081321),
(1.1685,1.3235,0.998302885095338),(1.172,1.3134,1.00745516824501),
(1.1658,1.3104,1.00553041729512),(1.1567,1.3008,1.00573267625465),
(1.1411,1.2769,1.00462125778582),(1.1439,1.2752,1.00411687920474),
(1.1623,1.2845,1.01698362656361),(1.1601,1.2963,1.03220478943022),
(1.1553,1.2924,1.03167234086454),(1.163,1.3068,1.03359173126615),
(1.175,1.3077,1.04307916970898),(1.1782,1.3159,1.03626943005181)
]
xor_outputs2 = [(2,),(2,),(1,),(1,),(3,),(0,),(2,),(1,),(3,),(0,),
(3,),(2,),(3,),(2,),(1,),(0,),(3,),(1,),(1,),(0,),(3,)
]
def eval_genomes(genomes, config): #function Used for training model
using the training set
for genome_id, genome in genomes:
genome.fitness = 84.0
net = neat.nn.RecurrentNetwork.create(genome, config)
for xi, xo in zip(xor_inputs, xor_outputs):
output = net.activate(xi)
genome.fitness -= (output[0] - xo[0]) ** 2 #Distance from
the correct output summed for all 84 inputs patterns
# Load configuration.
config = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,
neat.DefaultSpeciesSet, neat.DefaultStagnation,
'config-feedforward-research_nn')
# Create the population, which is the top-level object for a NEAT run.
p = neat.Population(config)
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
#p.add_reporter(neat.Checkpointer(100))
# Run until a solution is found.
winner = p.run(eval_genomes, 30000) # run for 30000 to test
generations
# Display the winning genome.
print('\nBest genome:\n{!s}'.format(winner))
# Make and show prediction on unseen data (test set) using winner NN's
genome.
print('\nOutput:')
winner_net = neat.nn.RecurrentNetwork.create(winner, config)
for xi, xo in zip(xor_inputs2, xor_outputs2):
output = winner_net.activate(xi)
print(" input {!r}, expected output {!r}, got {!r}".format(
xi, xo, output))
node_names = {-1: 'Input 1', -2: 'Input 2', -3: 'Input 3', 0:
'Prediction'}
visualize.draw_net(config, winner, True, node_names=node_names,
filename="winner-feedforward-research_nn.gv")
The config file:
[NEAT]
# changed
fitness_criterion = max
fitness_threshold = 75
pop_size = 500
reset_on_extinction = False
[DefaultGenome]
# node activation options
# changed
activation_default = sigmoid
activation_mutate_rate = 0.0
activation_options = sigmoid, square
# node aggregation options
aggregation_default = sum
aggregation_mutate_rate = 0.0
aggregation_options = sum
# node bias options
bias_init_mean = 0.0
bias_init_stdev = 1.0
bias_max_value = 30.0
bias_min_value = -30.0
bias_mutate_power = 0.5
bias_mutate_rate = 0.7
bias_replace_rate = 0.1
# genome compatibility options
compatibility_disjoint_coefficient = 1.0
compatibility_weight_coefficient = 0.5
# connection add/remove rates
conn_add_prob = 0.5
conn_delete_prob = 0.5
# connection enable options
# changed
enabled_default = True
enabled_mutate_rate = 0.04
feed_forward = false
initial_connection = full_direct
# node add/remove rates
# changed
node_add_prob = 0.5
node_delete_prob = 0.5
# network parameters
# changed
num_hidden = 0
num_inputs = 3
num_outputs = 1
# node response options
response_init_mean = 1.0
response_init_stdev = 0.0
response_max_value = 30.0
response_min_value = -30.0
response_mutate_power = 0.0
response_mutate_rate = 0.0
response_replace_rate = 0.0
# connection weight options
# changed
weight_init_mean = 0.0
weight_init_stdev = 1.0
weight_max_value = 30
weight_min_value = -30
weight_mutate_power = 0.5
weight_mutate_rate = 0.9
weight_replace_rate = 0.1
[DefaultSpeciesSet]
compatibility_threshold = 3.0
[DefaultStagnation]
species_fitness_func = max
max_stagnation = 20
species_elitism = 2
[DefaultReproduction]
elitism = 2
survival_threshold = 0.2
