What can Twitter tell us about stock investment?
Have you ever noticed a short-term fluctuation in the stock market caused by someone's words on Twitter? Have you ever regretted knowing it too late? Trader AI helps you spin the effect of Twitter-stock influence! It's a lightweight stock trading tool, powered by deep learning, and designed for individual traders. Now sit back and make some money!
As of January 2017, Twitter had 100 million daily user and 500 million tweets were sent per day1). With such an extensive reach, Twitter has become an invaluable communications tool for politicians looking to address their constituents on matters of policy and culture directly.
Trump’s blasts on Twitter about corporate moves that are “a disaster” or company policies that are “great” have reverberated on Wall Street, sending the shares of his tweet targets plummeting or soaring. Due to his influence as the “President of the United States” and his frequent use of Twitter, the “Trump Effect” has become an important market moving force.
Active traders who need to continuously monitor the market conditions to track down the ideal times for entering and exiting trades.
Individual Stock Investors
Those who need the best real-time news, simple and straightforward advice to help them profit from the rapidly changing market condition.
We present to you Trader AI
Trader AI continuously monitors Donald Trump's Twitter feeds and captures the ones that may have an impact on the stock price. It also provides price change predictions and trading advice to help you make the best deal.
Everything you need to know is here. Trader AI integrates critical information to help you evaluate the investment conditions.
With the Trader AI suggestions, you can place an order faster while still holding the advantages in your hand.
Not sure if you want to follow Trader AI's suggestions? It has a simulator that carries out every trading suggestion and keeps track of its profit and loss. There are risks, but we try to help you invest confidently with us.
I initiated the project and led a team of 5 to bring the concept to life from research to implementation. As the project manager, I defined the project scope, led the design and development process.
As the only designer in our team, I designed the entire product from low-fi to hi-fi deliverables and worked closely with the developers on implementation.
After the deployment of our original concept, I refined the product design individually focusing on improving the user experience. I carried out a series of user-centered design activities including user research, UX design, usability evaluation, and iteration.
In the United States, Twitter has exploded in popularity among political figures. It has become a necessary part of public life, with the sitting president and nearly every member of Congress actively participating.
The 45th President of the United States of America, Donald J. Trump, has used Twitter to address the public on a wide variety of topics. Trump’s frequent use of Twitter in the addressing of individual companies, countries and future U.S. policy towards commerce has certainly made his Twitter feed an important market moving force.
It appears that Trump’s tweets are similar to other types of media coverage in that they may focus attention over a trading day, but don’t carry much long-term weight for the markets. Jeffrey Born and colleagues at Northern University examined Trump's tweets between his election and inauguration when he tweeted about ten different companies. They found that positive or negative mentions may drive the stock price up or down during the trading day, but the impact would be gone three to five days later2).
Why did we choose to work with Donald Trump's tweets?
From the morning of his inauguration through its one-year anniversary, Trump has sent out 2,568 tweets, which is just a bit more than seven tweets per day3). This ensures the system built on Donald Trump’s twitter feeds gets enough training data, and will likely provide valuable information more frequently.
I led a team to explore the relationship between the influential Twitter feeds and the stock price movement. Specifically, we wanted to experiment with the idea of forecasting stock price change through real-time Twitter data analysis with deep learning.
While the team was doing the literature review in topics of stock price prediction with tweets, I went on searching for existing systems in the same area.
We weren’t the first group of people who want to spin the effects of President Donald Trump's tweets. After the market research, I found that the impact of presidential tweets on the U.S. equities market has been an ongoing topic of discussion throughout the financial community. Here are some open-source systems I found that aim to gain profit through the "Trump Effect".
- All of the above solutions only recognize tweets when they directly mention company names.
- All decisions on buying and selling stocks are made automatically by the algorithm.
- These models aren't able to capture the patterns besides sentiment.
- These systems all utilize sentiment analysis to determine whether the message was positive or negative towards the companies mentioned. The mainstream sentiment model computes a sentiment score by adding up the scores of each word in the sentence. But in some cases, the added score doesn't convey the meaning of the word sequence.
I consulted a few friends of mine who have invested in the stock market. They expressed their doubts, and there were some similar concerns:
“it’s really cool, but it’s not useful to me”
The main reason a stock investor wouldn’t use the auto trading bot is they think the risk is too high. In all those systems, the trading decisions are made by the AI without human interventions. Individual investors would want to have control over their investment. It’s hard for them to leave everything to an AI bot.
So...Why not give users the information they need to make their own decisions?
We came up with a solution where we introduced user interface. The AI captures influential tweets and presents the source, analysis, and suggestions to aid users' trading decisions. This solution focuses on delivering timely information and providing wise suggestions based on large data.
"No man is better than a machine, and no machine is better than a man with a machine
- Paul Tudor Jones
Stepstone - the previous work
It was a long journey to bring our concept to life. In the previous attempt, my team developed a deep learning model which became the foundation of our application. We then built a web application to present our findings and provide live suggestions for the individual traders.
Our innovation was using the end-to-end differentiable neural architecture to model the tweet-stock relationship. It extracts features from tweets using the long short-term memory network (LSTM), and feed them into fully-connected neural networks. LSTM yields state-of-the-art performance on many natural language processing (NLP) tasks because it has the ability to capture the long-term relationship in a word sequence.
The web application reveals the influence of Donald Trump’s tweets on the S&P 500 index between 2015 - 2018. Moreover, it predicts the stock price movement from real-time Twitter feeds and provides useful trade recommendations.
Investigate the history stock price
Stock price forcasting with tweets
Zheru Jiang (Design), Binghong Chen (Machine Learning), Jiahao Cui (back-end), Yincheng Zhao (back-end), Yujia Wang (front-end), Chu Han (data-processing)
Wharton Research Data Services (WRDS) database, Trump Twitter Archive
After we finished the deployment, I showed this to some individual stock traders. When asked "would you like to use it to help you in stock trading?", many of them gave similar comments:
The Problems Identified
A lot of unprofessional stock investors use mobile devices for stock trading. It’s unrealistic for them to keep the web application open all day to receive timely updates.
Challenges of timely response
Users will still need to manually place an order after they’ve been prompted to make a transaction. Sometimes they seek additional evidence and advice to confirm their assumptions. That requires jumping between devices and apps, taking extra time and may cause them to lose their advantage in the stock market.
Stock trading is such a problem of immediate concern because the money is real. Although we've had some improvements in the core algorithm and usability, the product design was not meeting the expectations of our users. To design a product that brings more value (money) to users, I put in more efforts to understanding users' needs.
I conducted semi-structured interviews with six people in my network. They were all unprofessional stock investors and had at least 1.5 years of experience in the stock investment. I framed my questions to focus on understanding their needs, pain points, and attitudes towards the tool.
The two types of stock traders
After the interviews, I realized that not all individual stock investors are excited about the type of service we promised.
The first type of investors views stocks as financial products. They would put their money in funds, or only in the companies that are growing steadily. They tend to hold on to them for long term, and profit from the economic growth or the company's growth. The short-term fluctuation in the stock market doesn't significantly affect their trading decisions. This type of stock investors typically doesn't want to spend a lot of time on stock trading. Their trading frequency ranges from once per month to once every six months.
The other type of stock investors, however, expressed keen interest in finding the "ideal time" for entering or exiting trades. They are actively seeking mispriced stocks, collecting information to help them "predict the future". They profit from the price difference. As a result, timing is critical in their strategies. This type of stock investors trade at a higher frequency, sometimes more than 10 times a day.
It became clear to me that I should design for the second type of stock traders. The features of our tool meet their needs for time-sensitive information. And the short-term predictions are very appealing to them.
I then developed a persona of my primary users - the "opportunist". The persona will help me maintain a consistent understanding of users' goals and needs throughout the process.
How can I help my target users achieve their goals? I reviewed my interviews with these users and wrote down the key takeaways on sticker notes. I then grouped them by affinity, trying to figure out the opportunities for the product design.
When provided with the prediction and suggestions from AI, users typically need to review the past price movement and relevant news to help them evaluate the investment condition. We can improve this experience by creating a simpler journey of collecting pieces of information.
Many of our users mentioned the importance of acting faster than the market. There's usually a short time window before things spread out. Our system needs to direct users to the transaction page within that time window.
Their biggest concern
Additionally, I discovered the biggest concern of my users. The common knowledge about stocks is that it's unpredictable. Users question the credibility of trader AI suggestions. An urgent user need is to see substantial evidence to confirm that Trader AI is providing helpful suggestions.
Understanding the user flow
Many opportunist traders make "information-driven" investment, which means their decisions are often driven by news and occasional information, such as a twitter feed. Now that I have a handful of knowledge about the users' needs and goals, I still lack the insights of where my product would fit in users' workflow. I decided to let users be the expert and learn from their process. So I gave users a prompt and observed them collecting information before making a trading decision. During the process, I asked them to explain their choices and express their feeling.
Then I labeled the negative experience to inform future product design choices.
Based on the research insights, I came up with a set of core features.
Machines are better than human in processing a large amount of data. The AI automatically monitors tweets and determines which tweet may affect the stock market.
Users can use some advice when it comes to figuring out the details of orders such as buy-in and sell price, the number of shares, stop win/loss point. The model can provide some useful reference learned from the historical data.
To help users outrun the market, we need to provide some quick and easy access to the order placement. If we ask users to connect their accounts of other stock trading apps, such as Robinhood, we can help them prepare orders with Trader AI and send it directly to their trading accounts.
Stock price prediction has been a difficult problem to tackle to this day. It’s very common that users will be skeptical about the performance of our system. While the prediction and trading suggestion can’t be correct every time, it’s worth trying if the long-term return is satisfying.
The solution I came up to address users' primary concern is a trading simulator. It will start with a fixed value and execute all the transaction suggestions. Users can have access to all the numbers they care about and compare trader AI's performance with other indexes.
The trading simulator makes the profit, loss, past suggestions transparent to users. It demonstrates the power of our system and can help users gain confidence when investing with us.
Although the process to arrive at this solution was complex, the product itself was straightforward. It is designed to help stock investors take advantage of the Twitter-stock effect in trading, and not to become a competitor of trading tools.
Sketching with pen&paper
The trending tweets and affected stocks
Prepare an order with Trader AI
Because most of the price fluctuations caused by Twitter feeds are short-term effects, Trader AI only gives trading suggestions as "Bracket orders". It aims to limit your loss and lock in a profit by "bracketing" each order with two opposite-side orders. For example, a buy order is bracketed by a high-side sell limit order and a low-side sell stop order.
The Trader AI Simulator
Evaluation & Iteration
At this point, I was proud of how long the way we've walked to arrive here and I loved how simple the final solution became. What do our users think of them now? With the ultimate goal being making products that users love, I conducted 5 evaluation sessions with my users.
The evaluation process
Before I invited my users, I designed a set of tasks and created clickable protoypes correspondingly. I also prepared a SUS survey to collect their feedback at the end.
- Set up the context
I welcomed my users, briefly introduced the project and walked them through the testing process.
- Present the prototype
I let users play with the prototype, asking them how would they complete the tasks. I encouraged them to think aloud while I took notes of their comments.
- Ask additional questions
Now it's time to ask follow up questions concerning the issues that occurred in the test. I also asked users if they had other questions that I can help answer.
- Give out the survey and thank my users
Lastly, I ask users to fill out the SUS survey and thanked them for participating.
Tasks and questions
- Identify the trending tweet, and which stocks it may have an impact on.
- Check out more details about Alibaba stock.
- Submit an order of Alibaba (the order details have been auto-filled).
- Add stock to your watchlist, then remove it.
- Visit the simulator page. Does everything make sense to you?
The evaluation result
In the end, I was happy to see all the participants found the tool to be useful to them. They still hold doubts about the effectiveness of trader AI suggestions (who doesn't?), but they were excited to have the tool as their reference. Additionally, some of them expressed strong interest in the trading simulator. If it allows them to customize the simulating strategy, according to my users, it would be a great standalone product for them to experiment trading strategies.
During these evaluation sessions, I also discovered some usability problems. I then iterated my design to solve those issues.
Problem - Users confuse the daily price change with the predicted change
The biggest issue discovered was that users tend to mistake the meaning of two price changes because the price prediction was put in a place where they are used to seeing "price change up to now".
Problem - How to prepare an order?
Problem - Users didn't expect to place the order from here
The voice of users
I initiated this project with a set of research objectives. The application was an interesting extension of our research. If it wasn’t for user’s participation, we wouldn’t have made it so far. We are more likely to create products that people love, if we keep users involved throughout the process.
Think big and start small
Every innovator had a big dream when they started. It was my first time leading a team to realize a concept from the initial research to development. We may not have fully achieved our vision, and some part of it is still an optimistic outlook. But each step forward is worth celebrating.
Potential Next Steps
Flesh out the simulating strategy
The simulator is a great way to demonstrate the power of our system. Some users were curious about the details of our simulating strategy, such as any stop winning/losing strategy, or the capability of user adjustments. It would be another interesting conversation in the future.
Donald Trump isn't the first and only politician to utilize Twitter for public voicing. Twitter is a popular media with many politicians and entrepreneurs actively participating. Its influence on the stock market is worth further exploring. Our next step could be to integrate more influential people's tweets, spanning other industries and regions.
Combine the web and mobile functionalities
The web application and the redesigned mobile app focus on meeting different goals. One of them better facilitates data analysis, and the other one aims to prompt better trading decisions. We can potentially combine their functionalities to enable more complex use case scenarios.