University of Maryland

Back to the Future: Exploring Multiple Futures in the Stock Market

A new tool called TimeFork uses interactive graphics and state-of-the-art statistical models to allow traders to investigate multiple parallel universes for predicting the stock market.

By Sriram Karthik Badam and Niklas Elmqvist

We all encounter time-series data—information that changes over time—in one form or the other in our everyday life: surveys or statistics reported in news articles, usage reports about electricity consumption from our energy company, or even the performance of the stocks we invest in. Predicting the future is a crucial task in these scenarios. For example, accurately predicting stock prices can increase profits from your investments. In a research paper published at the prestigious ACM CHI conference, we describe our new prediction technique called TimeFork that helps both experts and novice traders predict the future through the help of computer models. Our work shows that TimeFork leads to higher profits from stock trading compared to when analysts predict the future by themselves (without the models) or when the computer automatically predicts the future in isolation.

 Stock market data (stock price along with a grey band representing past variations) with two possible future trends, light blue versus. orange, generated by two different models. Light blue considers the past data of a stock and orange also considers the past trends of other stocks.

Stock market data (stock price along with a grey band representing past variations) with two possible future trends, light blue versus. orange, generated by two different models. Light blue considers the past data of a stock and orange also considers the past trends of other stocks.

Seeing the Future Ain’t Easy!

Current approaches for computer-aided prediction are based on learning from past behavior (Figure 1). However, this method cannot fully account for external factors that influence these trends. Assume that strong legislation banning fracking was just enacted; this legislation would certainly affect the oil stock prices, but no model could predict this behavior from market data alone. Similarly, Tesla’s recent unveiling (March 31, 2016) of Model 3 (a more affordable electric car for the masses) and the overflow of pre-orders (more than 300,000) have improved its stock price (link). Such external factors and event are often only processed by human analysts and not captured even in the most complex mathematical models. In other words, combining the knowledge of these human analysts with the mathematical models can be a promising new approach towards more powerful prediction.

Our TimeFork technique is based on this idea by helping analysts have a conversation with the computer model to combine the best of them both. The goal is to utilize the synergy between man and machine to achieve more accurate predictions than either computer-only or human-only alternatives (situations where a person or a computer model judges the future).

Combining the Best of Humans and Machines

TimeFork is a visual analytics approach towards prediction. The field of visual analytics combines advanced computer models with interactive visual pictures of the data to help the human understand. Using visual analytics in TimeFork allows analysts to have a conversation (Figure 2) with the computer, in which (1) the computer first shows multiple possible predictions using a computer model, after which (2) the analyst then provides their own opinion about the future (e.g., Apple’s stock might increase) based on their knowledge from news articles and social media, and finally (3) the computer takes the analyst’s assessment of the future into account and updates its predictions for the rest of the data.

By repeating this process, analysts can understand the effects of their guesses and come up with predictions that balance both human and machine “intelligence”.  In essence, the method allows the analyst to “fork time” and investigate “what-if” scenarios, i.e. multiple possible futures in parallel universes, to borrow terms from popular science fiction stories.

How TimeFork works. Far Left: Predictions are visualized (light blue). Left: User interacts. Right: Predictions globally revised on-the-fly (orange). Far Right: The user continues the interaction to come up with the best overall predictions.

How TimeFork works. Far Left: Predictions are visualized (light blue). Left: User interacts. Right: Predictions globally revised on-the-fly (orange). Far Right: The user continues the interaction to come up with the best overall predictions.

Better Stock Market Gains

We evaluated the TimeFork technique on real stock market data from 2014 and 2015 using student volunteers from the University of Maryland. Participants made investments in a simulation of the market using virtual money of $100,000. In these experiments, three preliminary trends emerged:

  1. Using TimeFork nearly doubled monetary gains for their investments (as compared to either manual human prediction, or purely computer-generated prediction).
  2. Participants using TimeFork who interacted more with the program made more money.
  3. Participants interacting with an intelligent prediction model made larger profits than those interacting with alternative methods.

The Future of Predicting the Future is Exciting!

The success of TimeFork shows great promise for time-series prediction and visual analytics in general. For example, TimeFork‘s workflow shows one way to enable analysts to directly “collaborate” with computer models as if they were another human. Furthermore, TimeFork enables the best of both worlds when we merging computer-generated predictions with human intuition. This suggests that future research and development should aim to refine the synergy of human and machine.

More Information

The TimeFork paper is accepted for publication at the ACM CHI 2016 conference. Here is the citation:

Sriram Karthik Badam, Jieqiong Zhao, Shivalik Sen, Niklas Elmqvist, David S. Ebert. TimeFork: Interactive Prediction of Time Series. In Proceedings of the ACM Conference on Human Factors in Computing Systems, to appear, 2016.

Additional resources:

Visualization research at the HCIL has a long and proud history, going back to the founding of the lab in 1983. Today, visualization researchers at the HCIL conduct research in information visualization, visual analytics, and data science. Sriram Karthik Badam is a Ph.D. student in Computer Science at the University of Maryland, College Park and a member of the HCIL. Niklas Elmqvist is an associate professor of information studies at the University of Maryland, College Park and a member of the HCIL.