Robert Buss and Dr. Scott Condie, Economics Department
Introduction
In recent years, high frequency trading has increased trade volume significantly. High frequency traders employ various algorithms in order to maximize their individual profit. Supporters of high frequency trading claim that high frequency traders provide the market with more liquidity, while opponents assert that such trading strategies make the market more unstable. Because high frequency trading is a relatively new phenomenon, academic research has not yet thoroughly investigated and modeled how markets currently behave. Additional research will allow policy makers to better understand what types of behavior are occurring in the markets and asses what effects they have in order to adjust financial regulations as deemed necessary.
Our research models how traders, particularly low-latency high-frequency traders, react to the signal implicit in the execution of of a hidden order. In order to examine the nearly instantaneous reactions characteristic of these traders we used NASDAQ ITCH data, which are the most detailed data available. However, these data do not come in a standard format that allows them to be examined without considerable cleaning and processing. Additionally, the data is proprietary and access rights can be very expensive. As a result, there is virtually no academic literature that examines the market down to the nanosecond level, which is essential to observing high frequency trading. When published, our research will be the first model tested against nanosecond data. Beyond adding to the literature, we will be able to provide other researchers with programs to process these data into a format that can be easily used, facilitating further examination of these important, but virtually unused data.
Methodology
The NASDAQ uses a limit-order book to conduct trading. Traders can place orders specifying what stock to trade, how many shares to buy or sell, and what price they will trade at. Orders are available for trading until canceled. Orders are traded on using a simple priority system: best price, display status, time of submission. Orders with the best price are executed first, priority being given according to display status (a regular visible order has priority over hidden orders), and time of order submission. Traders earn a small make fee per stock, paid by NASDAQ, for providing liquidity. Similarly traders must pay a small take fee when they buy stocks and reduce liquidity.
Our model is an extension of the standard Glosten-Milgrom model with three types of traders. Market makers, both informed and uninformed, try to make money by providing liquidity to the market without revealing their information. Speculators attempt to profit by reducing liquidity without revealing their information. Finally, noise traders with no objective introduce uncertainty into the model. The game-theoretic model predicts that the informed market makers will use hidden orders inside the spread (between the best buying and selling price) when their information signals that prices should not change and use regular visible limit orders outside the spread when their information signals that the prices will change. Thus hidden orders serve as a signal to the rest of the market indicating that the spread will narrow and visible orders away from the spread signal that prices will change and the spread will widen.
We were able to test this hypothesis using a six month sample of NASDAQ Total-ITCH data. The data comes in binary zipped files; in order to read the files decode them using the NASDAQ-specific codes. This gives us message-data. There are 19 different types of messages in the data, some of which communicate information relevant to the market including adding, canceling, and executing limit orders. By correctly processing the messages we are able to reconstruct the exact state of the market at any given nanosecond and see all changes on every stock traded on NASDAQ. This process is computationally intense and must be run in parallel on the Fulton Supercomputer. By simultaneously processing each days data on a hundred processors we can reconstruct the market and save key statistics to test the validity of this model. The data that we collect currently takes over 17 TB of storage space and cannot be handled using commercially available programs.
After processing the message data, we ran regressions on each days data for approximately 7000 stocks that had hidden orders executed. For each event (the execution of a hidden order), we looked at the change in the spread at various leads: the next order executed, 5ms, 50ms, and 500ms after the hidden orders execution. These regressions were manually coded up in Python in order to run on the Fulton supercomputer where our data are housed. By examining what lagged effect, if any, that the execution of a hidden order has on the spread we were able to test if the markets behave in accordance to the model.
Results
Preliminary analysis supported the model, however we discovered a few minor bugs in our code that required us to reprocess the data to insure that they were accurate. While doing so, we added interaction terms and control variables to the statistical model. These terms isolate the market’s reaction to a hidden order execution from any order’s execution. After removing the bugs and improving our statistical model, the results were more divided. I suggested that we divide the equities into quintiles based on the total number of messages for that equity. Stocks with a higher number of messages are more actively traded, and standard modeling assumptions are better met. As anticipated, the equities that are traded most frequently provide the most support for the model. Less frequently traded equities did not provide conclusive results for or against the model.
Discussion
Although the results were significant in favor of the model, the window of time where the market appears to act as the model predicts is too small; even the fastest algorithmic traders’ systems cannot receive information about the event and respond in such a short time. This result is possibly due to a type of order that automatically slides to react to changes in the market. This would be done on NASDAQ’s servers, traders would not have to observe the market event for their orders to change. This matches the reaction times we observed. Another possibility is that the observed reaction is not actually a reaction to the hidden order’s execution, but rather both the hidden order execution and the alleged market reaction are both driven by some variable not included in the model. Currently this possibility is being tested by looking at the market conditions before the hidden order was executed as well as after.
Conclusion
This research opens up a large dataset for analysis, and paves the way for using similar data from other markets. Although we have yet to finish show that our model is consistent with the data, we have made good progress to explaining the possible roles of hidden orders.