Predicting next 60 days using pytorch (LSTM)

I am trying to use pytorch to make predictions on time-series dataset.

1- First, I splitted the dataset into training and test.

Dataset

2- Then, I created the model.

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

        self.rnn = nn.LSTM(input_size=1,hidden_size=50,num_layers=3,batch_first=True,dropout=0.2)
        self.out = nn.Linear(50, 1)

    def forward(self, x):

        r_out, (h_n, h_c) = self.rnn(x, None)
        x = r_out[:,-1,:]                    #last hidden output!
        x = self.out(x)
        return x
    
model = Model()

if cuda:
    model.cuda()

3- After training the model, I made the validation step.

mse = mean_squared_error(dataset['y'].iloc[60:].values,predicted)
print("RMSE:", np.sqrt(mse))

# RMSE: 0.10680269716212222

Question: I would like to know, how can I use this model to predict the next 60 days ?

I read that is necessary to: Take the last prediction value(i.e predicted[-1]) from my predictions and add it to the last array of total_x, excluding the first timestep.

So, I tried this:

#array --> List
today = total_x[-1].reshape(-1).tolist()

# scaling last price
last_price = scaler.transform(infered['y_pred'][-1].reshape(-1, 1))

# adding last price to list.
today.append(last_price[0])

# Exclude first(0th index) element
today = today[1:]

#reshape
today = np.array(today).reshape(-1,60,1)

today = torch.Tensor(list(today))

#predict!
tomorrow = predict_with_pytorch(model, today)

#inverse transform.
tomorrow = scaler.inverse_transform(tomorrow)[0]

This Variable tomorrow is my tomorrow prediction, however, how to do this, for the next 60 days ?

1 Like

After you split your dataset into train and test, you need a create a data structure that looks like this:
Say, you have 10 rows in your dataset and you want to predict 1 value after looking at the previous 5 time steps.

1, 2, 3, 4, 5 -> 6
2, 3, 4, 5, 6 -> 7
3, 4, 5, 6, 7 -> 8
and so on…

So, X_train would be a numpy array:
[ [ 1, 2, 3, 4, 5],
[2 ,3, 4, 5, 6],
[3, 4, 5, 6, 7],
.
.
]

y_train would be a numpy array:
[5, 6, 7…]

Pass these into your RNN after converting this to tensors. Change shape as require by the rnn.

You can change your y_train to contain multiple numbers to predict multiple days into the future:
y_train:
[[5 , 6, 7],
[6, 7, 8],
[7, 8, 9]]

You’ll need to change the hidden_size argument in the rnn accordingly.

Hi @scarecrow21,

Thanks for your reply.

I have in mind, doing something like FbProphet, where I can make a prediction of the next 30 days looking to the whole past.

I am trying to do something like this:

def plot_results_multiple(predicted_data, true_data,length):
   plt.plot(scaler.inverse_transform(true_data.reshape(-1, 1))[length:])
   plt.plot(scaler.inverse_transform(np.array(predicted_data).reshape(-1, 1))[length:])
   plt.show()

#predict lenght consecutive values from a real one
def predict_sequences_multiple(model, firstValue,length):
  prediction_seqs = []
  curr_frame = firstValue
 
for i in range(length): 
    predicted = []        
    
    print(predict_with_pytorch(model, curr_frame[newaxis,:,:]))
    predicted.append(predict_with_pytorch(model, curr_frame[newaxis,:,:])[0,0])
    
    curr_frame = curr_frame[0:]
    curr_frame = np.insert(curr_frame[0:], i+1, predicted[-1], axis=0)
    
    prediction_seqs.append(predicted[-1])
    
return prediction_seqs

predict_length=60
predictions = predict_sequences_multiple(model, total_x[-1], predict_length)
print(scaler.inverse_transform(np.array(predictions).reshape(-1, 1)))
plot_results_multiple(predictions, total_y, predict_length)