Deep Learning Homework
This assignment calls for leveraging recurrent neural networks in the comparison of prediction models for bitcoin prices. Recurrent neural networks in general may have different structures and applications (https://www.cs.toronto.edu/~tingwuwang/rnn_tutorial.pdf). Also, one can consider the long short term memory (LSTM) recurrent neural network (RNN) as a methodology that "allows learning of long-term dependencies" (http://pages.cs.wisc.edu/~shavlik/cs638/lectureNotes/Long%20Short-Term%20Memory%20Networks.pdf). LSTM RNNs have been leveraged for predictive financial models (http://cs230.stanford.edu/projects_fall_2019/reports/26254244.pdf).
The data was drawn from the comma separated value (CSV) files. Then the data was processed through the models. The data analysis and data processing are recorded in the Jupyter notebooks. In addition, for different window sizes, losses were recorded by the student author (me, Franklin Bueno).
Window Size 10
5/5 [==============================] - 0s 13ms/step - loss: 0.0616 0.06157752126455307
Window Size 9
5/5 [==============================] - 0s 11ms/step - loss: 0.0607 0.06065821647644043
Window Size 8
6/6 [==============================] - 0s 5ms/step - loss: 0.0573 0.057325515896081924
Window Size 7
6/6 [==============================] - 0s 10ms/step - loss: 0.0584 0.05839494243264198
Window Size 6
6/6 [==============================] - 0s 7ms/step - loss: 0.0547 0.05467987060546875
Window Size 5
6/6 [==============================] - 0s 10ms/step - loss: 0.0490 0.04902857914566994
Window Size 4
6/6 [==============================] - 0s 5ms/step - loss: 0.0448 0.04480021074414253
Window Size 3
6/6 [==============================] - 0s 4ms/step - loss: 0.0392 0.039153698831796646
Window Size 2
6/6 [==============================] - 0s 5ms/step - loss: 0.0323 0.03228790685534477
Window Size 1
6/6 [==============================] - 0s 3ms/step - loss: 0.0267 0.026739951223134995
Window Size 10
5/5 [==============================] - 0s 5ms/step - loss: 0.1219 0.12194053828716278
Window Size 9
5/5 [==============================] - 0s 5ms/step - loss: 0.1434 0.14338549971580505
Window Size 8
6/6 [==============================] - 0s 5ms/step - loss: 0.1332 0.13317695260047913
Window Size 7
6/6 [==============================] - 0s 12ms/step - loss: 0.1230 0.12296931445598602
Window Size 6
6/6 [==============================] - 0s 4ms/step - loss: 0.1165 0.11649595201015472
Window Size 5
6/6 [==============================] - 0s 4ms/step - loss: 0.1211 0.12107281386852264
Window Size 4
6/6 [==============================] - 0s 4ms/step - loss: 0.1200 0.11998255550861359
Window Size 3
6/6 [==============================] - 0s 3ms/step - loss: 0.1138 0.11377636343240738
Window Size 2
6/6 [==============================] - 0s 5ms/step - loss: 0.1138 0.1137816309928894
Window Size 1
6/6 [==============================] - 0s 3ms/step - loss: 0.1117 0.11165456473827362
Which model has a lower loss? From the data above for loss, one can note that the closing price model has a lower loss.
Which model tracks the actual values better over time? In terms of tracking values over time, from the data in the graphs, one can see that the closing price model tracks the actual values better over time.
Which window size works best for the model? The data for loss and window size are provided above. After running each of the models for each window size from ten to one, the losses were recorded. In general, the loss was the least with the smallest window size (window size equal to one) for each model. It is worthy to note that with window size equal to one for each model, the loss was less for the student author than for the starter code (closing price model was loss 0.0267 for the student author and 0.0487 for the starter code, and fear and greed model was loss 0.1117 for the student author and 0.1911 for the starter code).
It is important to note that 99.9999% of this work comes from other sources, especially Instructor GS, Instructor AN, Instructor KS, the Tutor, Ms. LT, and the gold price prediction class exercise. The methodology for the 70/30 split of the data was taken from the gold price prediction class exercise. Also, the methodology for the training structure was taken from the gold price prediction class exercise.