Our team developed a novel stock-market prediction algorithm whose utility can be harnessed by way of a user-friendly, intuitive app. With the aim of both ameliorating the overwhelming barrier-of-entry for individuals interested in beginning down the investment path and bolstering/diminishing the confidence of seasoned stock-traders, this framework makes use of complex financial information within a seamless, inviting user-interface. Specifically, the workflow amalgamates historical data from technical analysis (i.e derivative crossings of moving averages) and semantic analyses (i.e press-release content) to provide a rich feature-set to a combination of support-vector-machine and lasso algorithms which then predict an individual company’s closing stock price for the next trading-day. As a proof-of-concept, we built a model for Amazon’s stock (NASDAQ: AMZN). Trained on technical and semantic data from 2014, our model was then tasked with predicting Amazon’s price for each day of 2015— an undertaking that significantly accounted for over 50% of the variance in stock-price fluctuations (r=.7, p=.000004) and achieved 83% accuracy in binary discretions (i.e whether the stock was up or down given the preceding set of information).
How unique was the idea behind the hack?
Our team combines extensive knowledge in financial analysis, machine learning algorithms, design, and cognitive neuroscience to create a unique attempt to predict stock market trends. While many have previously attempted to automate stock predictions, our hack is unique for two reasons in particular: (1) a synthesis of experience-based insights yielding data sets that are not typically incorporated in stock-prediction algorithms and (2) an intuitive and smooth user interface through which users will be able to interact with these data-driven insights in an informative and non-intimidating manner. Though these two attributes make us stand out from what is normally a very crowded space to develop novel solutions in, our favorite aspect of our project is the transparency we were able to achieve. When providing a prediction for upcoming stock prices, our app confidently displays the accuracy with which we’ve made similar positions in the past and provides specific insights as to the mechanisms contributing to this particular prediction. Thus, the end-user, in what we believe to be an unprecedented event, can easily produce a stock prediction, see what it would have been like if they followed through on this prediction in the past, and identify the key components behind this prediction simultaneously.
How technically challenging is the hack?
Our hack leverages the skills we have as cognitive neuroscience students interested in the stock market and app-development, namely our collective expertise in machine learning, financial analysis, and user-interface design. This project required proprietary information that a member of our group had acquired over a successful career as a day-trader; these insights would not necessarily to be obvious—even to those conversant in the financial sector. Quantifying this essentially guttural inclination was a substantial technological burden. Thus, the feature-collection aspect of this hack was particular challenging; our team had to build unique feature sets that drew on financial-time series data as well as conduct an in-house sentiment analysis that relied on both latent semantic analysis and linear discriminant analysis. Additionally, this project required a broad knowledge of machine learning algorithms capable of binary and continuous classification—a toolkit that the PhD students on our team utilize on a daily basis for “mind-reading” by way of combining neuroimaging (e.g fMRI) with machine learning. Lastly, the successful implementation of an end product mandated the development a robust user-interface mockup and nascent versions of the app for both iOS and Android mobile devices. This hack required a diverse and deep set of knowledge that few teams would be able to compile and execute over the course of a weekend.
How good was the implementation?
The foremost concern of our hack was the accuracy of our learning algorithm and retaining the ability to confidently and elegantly display its results. We have gone through numerous rounds of quality assurance to make sure our results accurately reflect the performance of our algorithms. These tests include classic machine learning sanity checks such as shuffling input data and using Gaussian noise as a feature set so as to demonstrate that the algorithm is not inherently biased in its predictions as a result of coding and data-organization error. Our results surprised even us—the ubiquitous success met by our prediction scheme is both reliable and powerful. Across more than a dozen companies, we could predict whether or not a stock would close up or down the following day (which would allow the end-user to act upon this information in the interim) with a range of 65% to 85%. Additionally, our workflow is capable of producing a continuous class label indicative of the magnitude of the stock’s movement. This class-label allowed us to create a follow-through simulation that made purchases and sales based on both the binary classification and the forecasted magnitude of the stock’s movement. That is, for each day preceding our training-set we could predict whether a particular stock would rise or fall, by what degree, and if that measure passed our user-provided risk aversion threshold. In so doing, we were able to create a simulation that showed the potential for capital growth of a user employing the services of YAMS on January 1st, 2015. Our simulations suggest that if someone were to begin using our algorithms for trading over the past year, an initial $5,000 would have grown into about $6,300. We feel as though this simulation will quickly stimulate the growth of a user-base as well as encourage revenue from potential advertisers (we can compare the capital growth when using a standard brokerage that charges $7.00 for each trade with that of RobinHood who tout a no-commission brokerage service).