Authors:
Vitaliy Milke
;
Cristina Luca
;
George Wilson
and
Arooj Fatima
Affiliation:
School of Computing and Information Science, Anglia Ruskin University, East Road, Cambridge, U.K.
Keyword(s):
Convolutional Neural Networks, Intraday Trading, Raw Financial Market Data, Machine Learning, Deep Learning, Supervised Learning.
Abstract:
This paper presents the use of Convolutional Neural Network (CNN) for finding patterns within intraday trading by being trained with raw Tick and other financial data. The network is specifically used to predict the probability of future movement at the intraday level of trading. The method of raw data pre-processing is evaluated and is critical to avoid errors that reduce the final accuracy of the model; for intraday trading, this includes a focus on the irregular Tick event rather than an arbitrary equal measure of interval time, such as a minute or a day. Training involves the use of a moving image window of 200 Ticks, where each increment of time is from 1 to 10 Ticks. For normalization (atypical for financial data) Tick intervals are capped at 20 milliseconds, Volumes are capped at 10 million, and Prices scaled over local extremes for each 200-Tick chart interval. The neural network was trained using the publicly accessible cloud computing GPU processors of Google Colaboratory.
An original methodology for selecting the training data was used which reduced the number of calculations by including only patterns close to the active movements of interest.
(More)