I like to study macro-economics and it is quite essential when reading stock market prices since they are heavily influenced by it. It is quite remarkable how a single statement from the Federal Reserve chairman, Jerome Powell and the entire stock price completely changes course. I usually spend time watching the Fed Minutes and also find interesting how several news outlets ask Jerome tricky questions to pull more information to expose hidden truths. The motivation for this particular project was to use Microsoft Stock prices and relate them to historical events that have impacted our economy to tell a compelling story about tech and economics. I also wanted to build some deep learning models to see if I can try to predict the movement of the stock price.

• Yahoo Finance

• Data Preprocessing
• Deep Learning Modeling
• Power BI
• App Deployment

• R (Deep Learning modeling)
• Power BI (Dashboard creation)
• HTML/CSS (website)
• GitHub Pages (web app deployment)

Long Short Term Memory - Prediction Plot

Gated Recurrent Unit - Predict Plot

Deep Learning Algorithms Results

• Activation Function Used: tanh
• Loss Function: Mean Squared Error
• Metric Used: Mean Squared Error
• Best Model: Gated Recurrent Unit
• Why use Mean Squared Error: Observe that the goal of this project was for our deep learning model to predict the closing price of the stock for that day. We were dealing with a regression problem and metrics such as F1-Score, Recall, and accuracy do not make sense in this scenario. Now, Mean Squared Error value represents the average squared difference between the predicted values and the actual values. Since we are taking the average squared difference, MSE can never be negative and the lower the MSE the more indication that the model performed a better fit to the data. Now if we take the square root of the mean squared error, we get the root mean squared error which takes the same units of the actual data. In short hand, its similar in metric but it is more interpretable and easier to compare to the actual data. For example, we can say for GRU model which had the best mean squared error compared to the two had a root mean squared error of about 9.63 thus we can say on average the predicted price was about $9.63 away from the actual price. This where we can take the standard deviation to make a confidence interval if we truly wanted to.

• The biggest lesson learned in this project was the difficulty in setting up a deep learning environment in R Studio. I did not know that it the end it uses python and same libraries such as tensorflow and keras. I would never work with deep learning models in R again since I already have working knowledge in python, my preferred programming language. However, R still reigns supreme when it comes down to statistical analysis and still conduct many of my hypothesis tests in R along some Linear/Logistic Regressions.
• Another major lessons learned was to properly chose the number of epochs (steps to train the model) since we can potentially over fit the model. Over fitting the model would imply a low bias but high variance meaning our testing set predictions we be far off for our actual values. We want the perfect balance between bias and variance.

  

Website Deployed

Video to gif tool