smohammadi96 / stock_prediction_lstm

comparison LSTM and GRU for stock prediction

stock prediction

Install dependencies

• pandas
• numpy
• maplotlib
• keras
• tensorflow>=1.18

dataset sample

Results

The plot below, shows the actual value of the close price in the training and testing phase and the actual value as specified The model has been able to predict the Close price well. The red color is the actual value of the test data, the orange color, The output value of the model and the blue color shows the value of the close price in the training process.

LSTM

GRU

First Header	Layers number	neurons number in each layer	Optimizer	Rmse without scale	Rmse with scale
GRU (tanh)	3	128-64-32	Adam	43.14	0.01
LSTM	2	128-64	Adam	57	0.02
GRU	3	128-64-32	Adam	59.74	0.02
LSTM (tanh)	2	128-64	Adam	490	0.19
LSTM	2	128-32	Adam	53.02	0.02
GRU	2	128-64	Adam	148	0.05

The best result that has the lowest MSE is related to the 3-layer GRU network with the number of neurons 64, 128 and 32, which have better results than LSTM. GRU and LSTM both use different ways to avoid vanishing gradient problem. The main difference between GRU and LSTM is that GRU has two gates that are reset and updated, but LSTM has three gates: input, output, and forget This is because GRU is less complicated and more suitable for small datasets. It is worth mentioning which exposes GRU full memory and hidden layers.