PyBrain neural network for stock prediction won't learn

Question

I am trying to code a neural network that can forecast some data. Therefore I use PyBrain for python. I figured out that a SupervisedDataset would be a good fit for this task. I took some stock data and put 5 values from it as input and the sixths as the target. Then I build a feed forward network with the buildNetwork function and trained it with the BackpropTrainer.

Anyway, the error won't get less. It is stuck at ~0.6 and it seems to oscillate around there. I tried to tweak the momentum and the learning rate but it didn't help. What am I doing wrong?

from pybrain.datasets import SupervisedDataSet
DS = SupervisedDataSet(5, 1)

DS.addSample((44.055, 44.54, 44.04, 43.975, 43.49), (42.04,))
DS.addSample((44.54, 44.04, 43.975, 43.49, 42.04), (42.6,))
DS.addSample((44.04, 43.975, 43.49, 42.04, 42.6), (42.46,))
DS.addSample((43.975, 43.49, 42.04, 42.6, 42.46), (41.405,))
DS.addSample((43.49, 42.04, 42.6, 42.46, 41.405), (42.385,))
DS.addSample((42.04, 42.6, 42.46, 41.405, 42.385), (42.655,))
DS.addSample((42.6, 42.46, 41.405, 42.385, 42.655), (41.53,))
DS.addSample((42.46, 41.405, 42.385, 42.655, 41.53), (40.09,))
DS.addSample((41.405, 42.385, 42.655, 41.53, 40.09), (39.8,))
DS.addSample((42.385, 42.655, 41.53, 40.09, 39.8), (40.2,))
DS.addSample((42.655, 41.53, 40.09, 39.8, 40.2), (39.915,))
DS.addSample((41.53, 40.09, 39.8, 40.2, 39.915), (40.21,))
DS.addSample((40.09, 39.8, 40.2, 39.915, 40.21), (40.34,))
DS.addSample((39.8, 40.2, 39.915, 40.21, 40.34), (41.195,))
DS.addSample((40.2, 39.915, 40.21, 40.34, 41.195), (41.595,))
DS.addSample((39.915, 40.21, 40.34, 41.195, 41.595), (41.975,))
DS.addSample((40.21, 40.34, 41.195, 41.595, 41.975), (42.045,))
DS.addSample((40.34, 41.195, 41.595, 41.975, 42.045), (40.13,))
DS.addSample((41.195, 41.595, 41.975, 42.045, 40.13), (38.99,))
DS.addSample((41.595, 41.975, 42.045, 40.13, 38.99), (39.81,))
DS.addSample((41.975, 42.045, 40.13, 38.99, 39.81), (40.23,))
DS.addSample((42.045, 40.13, 38.99, 39.81, 40.23), (40.47,))
DS.addSample((40.13, 38.99, 39.81, 40.23, 40.47), (40.45,))
DS.addSample((38.99, 39.81, 40.23, 40.47, 40.45), (40.01,))
DS.addSample((39.81, 40.23, 40.47, 40.45, 40.01), (40.23,))
DS.addSample((40.23, 40.47, 40.45, 40.01, 40.23), (40.2,))
DS.addSample((40.47, 40.45, 40.01, 40.23, 40.2), (41.605,))
DS.addSample((40.45, 40.01, 40.23, 40.2, 41.605), (42.1,))
DS.addSample((40.01, 40.23, 40.2, 41.605, 42.1), (42.135,))
DS.addSample((40.23, 40.2, 41.605, 42.1, 42.135), (41.95,))
DS.addSample((40.2, 41.605, 42.1, 42.135, 41.95), (41.145,))
DS.addSample((41.605, 42.1, 42.135, 41.95, 41.145), (40.635,))
DS.addSample((42.1, 42.135, 41.95, 41.145, 40.635), (41.25,))
DS.addSample((42.135, 41.95, 41.145, 40.635, 41.25), (41.19,))
DS.addSample((41.95, 41.145, 40.635, 41.25, 41.19), (42.065,))
DS.addSample((41.145, 40.635, 41.25, 41.19, 42.065), (42.025,))
DS.addSample((40.635, 41.25, 41.19, 42.065, 42.025), (42.09,))
DS.addSample((41.25, 41.19, 42.065, 42.025, 42.09), (41.79,))
DS.addSample((41.19, 42.065, 42.025, 42.09, 41.79), (43.11,))


from pybrain.tools.shortcuts import buildNetwork
FNN = buildNetwork(DS.indim, 15, DS.outdim, bias=True)

from pybrain.supervised.trainers import BackpropTrainer
TRAINER = BackpropTrainer(FNN, dataset=DS, learningrate = 0.005, \
    momentum=0.1, verbose=True)

for i in range(1000):
    TRAINER.train()

Edit: Some of the comments doubted that those data would fit for a neural network in general. Therefore I did the same net in MATLAB and it worked just fine. After 11 training epochs the error was less then 0.002.

same net and data in MATLAB

Furthermore I tried to use the SupervisedDataset from PyBrain but this wouldn't work as well. I am out of ideas now.

While I don't have a definitive answer to your question, I'd like to point out that you aren't, necessarily, doing anything *wrong*. Your data might not simply have information there to predict - if you *could* use a FFNN to predict your particular stock market in order to turn a profit, somebody else would already have done so. If you're doubting your code, my advice is to test your network on easier data first, and go from there — loopbackbee, Sep 15 '14 at 10:21
Also, you don't mention *which* five values you're using as training data. Assuming it's the 5 consecutive values before the one you want to predict, note that it's unusual to feed the NN raw data. Most work that I'm aware of works on processed data and higher level statistics — loopbackbee, Sep 15 '14 at 10:27
Thanks. Thats just for learning purposes only. In fact i am just trying to get a better result for my prediction than using the value from the day before. I will trying to test the network with easier data. — Gizmo, Sep 15 '14 at 10:30
You could also vary the number of previous days you use as input. I'm not sure if NN are the right tool here to predict a timeseries with absolute values. You might want try other targets as well. Maybe just Plus and Minus to predict if the stock will rise or fall. — Max Linke, Sep 15 '14 at 10:56
But shouldn't the network at least adjust to the given samples? Despite of being useful for other data the network should be able to learn this samples and minimize the error or am i wrong? The error really won't get better. — Gizmo, Sep 15 '14 at 11:09
Here is a book explaining why your error will probably remain high: http://www.amazon.com/Random-Walk-Down-Wall-Street/dp/0393330338 — John Zwinck, Sep 15 '14 at 13:01
You could, if you insist on using stock data, use stocks (such as six from the same sector) that historically tend to go up and down together. But people have been trying to get basic neural networks to predict stock prices for several decades now, I recall reading about it in the early 90's. If it worked, it would be widespread, which would mean it would stop working. Not even the neural networks in human brains are able to do it consistently, and they're much bigger than your software neural network. — rossdavidh, Sep 15 '14 at 13:22

Gizmo · Accepted Answer · 2018-10-09T08:30:48.870

I found a solution. Turned out the stock data had to be normalized first. So i wrote this function:

def normalization(data, new_max, new_min):
    old_max = 0
    old_min = 0

    # Finde altes Max- und Minimum
    for i in range(len(data)):
        if old_max < data[i]:
            old_max = data[i]
        elif old_min > data[i]:
            old_min = data[i]

    old_range = (old_max - old_min)

    for i in range(len(data)):
        if old_range == 0:
            data[i] = new_min
        else:
            new_range = (new_max - new_min)
            data[i] = (((data[i] - old_min) * new_range) / old_range) + new_min

I scaled the data between 0 and 1 and voilà - the network would finally learn.

PyBrain neural network for stock prediction won't learn

1 Answers1