stock prediction : GRU model predicting same given values instead of future stock price

Question

i was just testing this model from kaggle post this model suppose to predict 1 day ahead from given set of last stocks. After tweaking few parameters i got surprisingly good result, as you can see. mean squared error was 5.193.so overall it looks good at predicting future stocks right? well it turned out to be horrible when i take a look closely on the results.

as you can see that this model is predicting last value of the given stocks which is our current last stock.
so i did adjusted predictions to one step back.. so now you can clearly see that model is predicting one step backward or last stock prise instead of future stock predictions.

This is my training data

# So for each element of training set, we have 30 previous training set elements 
X_train = []
y_train = []

previous = 30

for i in range(previous,len(training_set_scaled)):
    X_train.append(training_set_scaled[i-previous:i,0])
    y_train.append(training_set_scaled[i,0])
X_train, y_train = np.array(X_train), np.array(y_train)


print(X_train[-1],y_train[-1])

This is my model

# The GRU architecture
regressorGRU = Sequential()
# First GRU layer with Dropout regularisation
regressorGRU.add(GRU(units=50, return_sequences=True, input_shape=(X_train.shape[1],1)))
regressorGRU.add(Dropout(0.2))
# Second GRU layer
regressorGRU.add(GRU(units=50, return_sequences=True))
regressorGRU.add(Dropout(0.2))
# Third GRU layer
regressorGRU.add(GRU(units=50, return_sequences=True))
regressorGRU.add(Dropout(0.2))
# Fourth GRU layer
regressorGRU.add(GRU(units=50))
regressorGRU.add(Dropout(0.2))
# The output layer
regressorGRU.add(Dense(units=1))

# Compiling the RNN
regressorGRU.compile(optimizer='adam',loss='mean_squared_error')
# Fitting to the training set
regressorGRU.fit(X_train,y_train,epochs=50,batch_size=32)

And here is my full code, you also able to run this code at google colab.

so my question is what is the reason behind it? what am i doing wrong any suggestions?

Maybellemayberry

How did you handle the problem.? – Plectrum

i am still trying to find the answer but from this post you can see some updates in the Sequence Generation part, but i have not try it yet. – Maybellemayberry

Answer 1

It is a well-known issue with regression actually. Since the task of the regressor is to minimize error, it secures it task by choosing the closest value from the features you input to the regressor. It becomes the case especially in the time-series problems.

1) Never give unprocessed closing value that you want your model to predict, especially in the time-series regression problems. More generally, never give a feature that gives some direct numerical intuition to a regressor about what the label might be.

2)If you are not sure whether the model just replicates like your case, be sure to plot the original test set and your prediction all together to visually analize the situation. Moreover, if you can, do a simulation of your model on the real-time data to observe whether your model predicts with the same performance.

3)I’d recommend you to apply binary classification rather than regression.

I’ve been intensely working on financial signal prediction for nearly a year, do not hesitate to ask more.

Have fun.

Sabian

hey thanks for this detailed replay, now as per following your answer i am going with the binary classification. let's see how it performs. Edit: i am also thinking to do multi-sequence prediction is that a good idea? – Maybellemayberry

You should provide not just the last sample, a time-series window that has last n samples as just one input. For example; if you want to predict the up/down fluctuation at time t+1, provide features of t, t-1, t-2,....,t-29 as a single sample to the model for each point for a larger pattern. If that is what you mean, totally go for it. But if you mean that you want to predict not just the t+1, but t+2 and so, that is dangerous; you should come with something clever with that. – Sabian

@UgurMULUK, I am also dealing with a similar problem, can you take a look here #52252942 – Plectrum

Answer 2

In my opinion this is happening because your loss function here is mean squared error. The best way to minimize this error is by using the previous value as the estimate for the current step (especially if the steps are nearby).

You will start seeing the limitations of your model clearly if you try to predict the delta between samples.

I am currently experiencing similar issues trying to predict using time series. I haven't found a solution yet, but am working on extracting features as others pointed out.

Gist