Predict next month's price/oz. of gold based on the price over the preceding five months, using sequence generation via ANNs with LSTM neurons

This is the result of some experimentation I did involving deep learning + sequence/time-series prediction using recurrent neural networks in a finance context.

Time-series prediction applied to stocks carries with it a certain set of difficulties, mainly due to changes in the value of the companies behind those stocks which can affect a stock's performance to a significant degree. Just analyzing a price graph won't really give you a totally accurate picture of stock market movements.

Since gold is a commodity whose supply and consumer demand are relatively stable, price fluctuations are typically generated by large institutions like bentral banks, in response to general market movements. Since these are (somewhat) cyclical in nature, there might be some discernable pattern that a neural network can actually learn and reproduce.

And in any case, I'm pretty sure everyone has done stocks at this point. Figured I'd try my hand at doing something a little different.

Yes, this turned out to work pretty well. When trained and tested on train/test subsets of the Bundesbank gold price dataset from Jan 1st, 1950 - Jun 6th, 2014 (included in this repo), the model achieved an average error of about 6 dollars. You can run the test yourself, too. Just clone the repo, cd into it, and run python test_model.py.

Here's the resulting normalized graph generated by the test script (x-axis is time, and y-axis is the price/oz. of gold:

The blue line is the original dataset, the green is the training subset used to train the model, and the red is the generated predictions for the subset of the data not seen by the model.

Clone the repo, cd into the root, and run python predict.py with 5 system arguments, representing the gold price, in dollars, on the first day of each of the five preceding months, in chronological order.

Example: "python predict.py 1243.068 1298.713 1336.560 1299.175 1288.913"

This command will print out the predicted price on the first day of the following month to the console.

##Resources + further reading Check out this tutorial to see the base code from which I built the model. Also click here for the github repo which has the data I used to train and test.

And click here to access an article on gold prices and what affects them.