LSTM to predict stock price and judge decisions

Inspiration

As a team, we were looking for the most efficient ways to do a time series forecast in order to predict stocks. We researched several research papers and drew much of our inspirations from work that used reinforcement learning to better predict the fluctuations of stock prices in the future.

What it does

Given a series of inputs, it uses a LSTM model to forecast future stock prices and gives an output to sell, buy, or hold on shares of stock. It is different than other models because it incorporates open_today open_prev high_prev close_prev low_prev adj_close_prev volume features, and most models only base their predictions off the open or close price.

How I built it

We built it using an LSTM model that predicted future values of stock prices and coded an algorithm that would use those prices to determine whether to sell, buy, or hold on those shares of stock.

Challenges I ran into

In the beginning, we didn’t know too much about reinforcement learning and had many difficulties on the formatting of the input, output, and submission. We were stuck for a long time trying to figure out the bugs that were present in the starter code. From there, we had to improve a solution given the dependencies provided. Another challenge that we encountered was the fact that the data given had extremely small values, so we ran into challenges with training and accounting for that data.

Accomplishments that I'm proud of

We are very proud of the extent to which we were able to improve the stock predictions and with this being the first time much of our team has implemented machine learning models into a real world scenario, the level of improvement and depth of understanding we gained was invaluable. We were initially hesitant about how much financial profit our model could really enable so being able to increase the value of the initial investment several fold was also very rewarding.

What I learned

As a team we gained a range of skills from our iterative development process. We got new insight into the process of transfer learning and how it can make innovation more efficient. In addition, we all had different levels of experience with machine learning techniques coming into the datathon so we helped each other hone our skills in data analysis and AI frameworks. Lastly, we all learned a lot about the scope of machine learning, how it can improve various facets of our financial decisions, and the various models that are still being optimized and tested to make AI/ML even more effective.

What's next for LSTM to predict stock price and judge decisions

I think next steps would include finding a better way to deal with the data given and creating better models to account for future forecast predictions and decisions on whether to buy, sell, or hold. We were originally trying to implement some of the techniques that the most recent research papers employed, but ran into several issues. If possible, our next steps would be to emulate some of their models and improve/configure to make a better model.