Time series forecasting is a supervised learning problem where we try to predict the current stock price of a company/business by looking into the stock prices of the past few years. Here we re-frame our data in a way so that we are able to use standard machine learning algorithms on it. Since we are trying to predict a real value here, it is a regression problem. I used LSTM model architecture (deep learning) for doing this project.

1. I used a data set named Lumber-futures.csv
2. The data set has 11378 observations and 9 features
3. The data set started with the most recent stock price data and ended with the oldest
4. After reading the csv file as a pandas dataframe, I reversed the order of the observations so that they start from the oldest and finish with the latest stock price, the dates range from 11/16/1972 to 2/2/2018
5. The column that is important for this kind of project is the closing stock price for a particular day, in this data set that column is called 'Last'
6. I plotted the data using just the 'Days' feature and the closing price 'Last' feature using matplotlib

1. Here the data 'Last' column is separated out as a new data frame and is converted to a numpy array
2. The data is then scaled (I have used MinMaxScaler, so the values will be within 0 and 1) (required step before feeding the data into an LSTM model)
3. The data is then divided into the training and test data sets (I used a 0.7 - 0.3 split)
4. Within the training data we need a x_train (features) and y_train (label), here we have the window size that determines how many previous days observations do we want the current stock price to look up to, I used a window size of 100, thus the observations, indices 0 to 99, will be put into our x_train and index 100 will be put into y_train
5. The same procedure is performed on the test data as well

1. I used a stacked LSTM (Long Short Term Memory) model architecture
2. Model is compiled using the default values and mean squared error loss is calculated
3. The model is trained with the data and the label with a batch_size of 64 and 100 epochs
4. Then the testing data is fed into the model for prediction
5. The root mean squared error is calculated on the differrence between what the model predicted and what the actual labels were
6. The error value is 1.866

• Finally, the training data, test data and the model's predictions are plotted using matplotlib