Using Big Data to Analyse Dynamics of Stock Price Discovery


Photo: Price response to Apple’s earnings news released at 4:30 PM October 2011 at one-minute frequency. Stock prices are plotted in dots (top panel) and in the bottom panel the blue bars represent buy orders and the red bars are sell orders that enter the NASDAQ trading platform. These charts show how fast prices respond to firm news. (Contributed by Charles Martineau)

When a company releases its quarterly earnings, investors want to know how long it takes for any unexpected news to be reflected in the company’s stock price.  “My research focus is on the dynamics of price response to news and how the news is incorporated into stock

“My research focus is on the dynamics of price response to news and how the news is incorporated into stock prices,” says Charles Martineau, a Doctoral Student at the Sauder School of Business at the University of British Columbia.

This question lies at the core of financial economics and market price efficiency and is often referred to as Price Discovery.  When a higher than anticipated value is announced, stock prices tend to increase, after the announcement. Martineau’s research focuses on how long that takes, and therefore when such news is fully incorporated into stock prices such that an investor can no longer trade on the news and earn a profit relative to less informed investors.

“When the news is released,” says Martineau, “do prices adjust automatically, or does it need the trading process to happen first for the news to be fully reflected in prices? If prices don’t respond to news instantly, how much time does it take? One second, 30 minutes, or a week?”

So why does it matter if it takes one second or a week for news to be incorporated into asset prices?

Martineau explains that if markets are not fast enough at incorporating unexpected news into stock prices then there are opportunities for some investors to buy or sell stocks at prices that are not reflective of the true stock value.

“As investors, we want prices to reflect public news as fast as possible,” says Martineau. “In my research, I find that it takes a few minutes to a few hours at most for stock prices to reflect earnings announcements. But despite fast price discovery, I found some unexpected findings that push our understanding of the functioning of financial markets to another level.”

He explains that this discrepancy is most at play during after-market hours, from 4 pm until 9:30 am, a period some warn should be reserved for sophisticated investors only. This is because firms wait to announce corporate news until after markets have closed, meaning stock prices during after-market hours may not be reflective of new announcements. Retail investors (i.e. the general public) are often advised to trade only during regular market hours (between 9:30 a.m. and 4 p.m.) because after this time there is no obligation to provide liquidity – a bid and ask price for stocks. Therefore sophisticated institutional traders may have an unfair advantage over retail traders who may not know the prices have not had time to adjust to recent announcements.

Martineau notes that his research shows that by the time markets open at 9:30 a.m., any earnings news is fully incorporated into stock prices. This work has involved working with some major data sets.

“To process this is very challenging and computationally demanding – for both storage and memory,” Martineau says, “and resulted in almost 25 TB of data. Without the resources provided by Compute Canada and WestGrid, my research would not be feasible.”

Martineau’s dataset involves all orders that enter the NASDAQ stock exchange trading platform, timestamped at the nanosecond. For one given day, there are approximately 300 million order messages, which over five years results in approximately 5 TB of compressed data. Martineau then reconstructs for each stock, in real-time, the limit order book, which is all the quantities available to buy or sell of a given stock within a price range.

“Python is my savior in combination with Hermes and Breezy,” says Martineau. “Doug Phillips at the University of Calgary, Belaid Moa from UVic, Roman Baranowski at UBC, Masao Fujinaga at the University of Alberta, and others provided tremendous help for getting me up and started with WestGrid.”

He explains that using Amazon Cloud or other powerful fee-for-service computers would have been too costly.

“WestGrid and Compute Canada provides researchers in finance with almost an unfair advantage relative to many U.S. schools. Compute Canada is one of the main reason I decided to stay in Canada when I accepted an assistant professor position at the University of Toronto.”

In terms of next steps, Martineau plans to revise his work on earnings announcements in combination with another big data source outside NASDAQ to verify that the results hold on multiple exchanges. He plans to turn to WestGrid and Compute Canada resources and support for this analysis too.

“I will use textual analysis and machine learning on all the New York Times and Wall Street Journal articles to construct measures of investor attention to different macroeconomic risks and use these measures to link with financial market asset price volatility. The data is not quite as large as the financial transaction data I have used in other studies, but it just goes much faster with WestGrid resources.”

Leave A Comment