Neural Networks in Limit Order Execution Time Prediction

This thesis proposes a convolutional long short-term memory neural network model for predicting limit order execution time. The prediction task was defined as a regression problem of how long it takes for a limit order with a certain side, price and quantity to be completed. Only full completions that took less than an hour to complete were considered. The objective of the thesis was to examine the performance of neural networks in the prediction task. The predictions were conducted on a 20-day data set of stocks of Apple, Google, Intel and Microsoft. Two neural network models were constructed to produce predictions: a multilayer perceptron model and a convolutional long short-term memory model. The results were compared against the results of a naïve model and a simple linear regression model. Two sets of input features were built using the message book and limit order book data: a nonsequential data set containing information of the order being initiated and the state of the market at the time of the initiation, and a sequential data set aiming to capture temporal dynamics within the market during the last hour from the initiation. The proposed methods outperformed the naïve and linear regression models. The improvement to the naïve model was substantial measured by both mean squared error and mean absolute error on all four stocks. In comparison to the linear regression model, the mean absolute error saw a significant improvement, whereas the mean squared error was improved considerably less. The convolutional long short-term memory model outperformed the multilayer perceptron model, indicating that temporal dynamics exist within the limit order book data that affect the execution time