Dynamic Wave Recognition in Financial Time Series with Neural Network

How to effectively recognize price movements in financial time series is a challenging task with the learning power of deep neural network in advanced quantitative analysis. It is well known that the common sequential data in stock markets composed of prices carrying hierarchical information including trend and wave in multi-scale, which directly reflect the overall direction of price movement. How to extract the wave pattern efficiently from sequential data of stock market under the premise of determinate trend direction will essentially affect the success or failure of investment activities. Stock price prediction as a form of financial time series analysis is considered complex, volatile, dynamic and chaotic with high uncertainty, non-linearity and strong dependency on empirical and theoretical knowledge. Comparing with the cross-sectional data, time series data contains complex time-domain features. Thus, it will increase the difficulty of prediction on stock price sequence. Since such a sequence usually has dynamic length, it is infeasible to train with traditional neural networks. Even if sequences are processed with equal length, due to its dynamic essence across time, it is impossible to directly compare and learn feature of waves in different scales with shift and distortion in time. In addition, the well-known sliding window approach with fixed length will disrupt the trend and discard valuable information, so that the consistency of intrinsic feature in limited time steps cannot be guaranteed.

The Influence of Sample Reconstruction on Stock Trend Prediction via NARX Neural Network

Through the established NARX neural network model, the sample data of stock AAPL in NASDAQ from 2006/01/01 to 2015/01/01 are utilized for training. The results show that under the same sampling frequency, with the increase of MA period, the trend of volatility becomes lower with obvious longer time delay, which will help to predict the trend of movement. In addition, through the use of reconstructed data containing the trend information as training sample, it has significantly reduced the prediction error, which is 16.29% lower than using daily training sample and 16.90% lower than using weekly training sample. The outputs directly reflect the probability of trend movement at every time point in stock price. It also improves the generalization ability of NARX model, so as to predict the stock trend change at a certain time. It has successfully estimated the possibility of buying and selling points, which provides the necessary theoretical basis on how to determine the stock trading points.

Related Publications

• Y. Wei and V. Chaudhary, “The Influence of Sample Reconstruction on Stock Trend Prediction via NARX Neural Network”, 14th IEEE International Conference on Machine Learning and Applications, ICMLA 2015, Miami, FL, USA, December 9-11, 2015, 2015, pp. 52–57. pdf

Fast Quantitative Analysis of Stock Trading Points in Dual Period of DMAC

We propose a novel volatility segmentation approach to detect effective trading points from 2679 stocks of NASDAQ. The buy and sell points are derived from dual periods of DMAC based on daily and weekly periods by estimating the amplitude and interval. The proposed approach is very accurate in that only 373 stocks (out of 2679) in NASDAQ have the average rate of profit of overall buy points higher than 3% and only 193 stocks (out of 2679) in NASDAQ have the average rate of stop-loss of overall sell points higher than 3%. The volatility segmentation approach reduces the uncertainty of stock estimation in single period DMAC. This approach, however, is very computationally intensive requiring 2382.82sfor evaluating buy points and 2688.53sfor evaluating sell points. A parallel implementation using 240 cores reduced the time to 16.01s and 13.50s, respectively.
 
Related Publications

• Y. Wei and V. Chaudhary, “Fast Quantitative Analysis of Stock Trading Points in Dual Period of DMAC”, 2015 IEEE First International Conference on Big Data Computing Service and Applications, 2015.pdf

Financial Forensic

Nowadays finance is inundated with a flurry of data that is increasing in complexity as well as in size. For example, the data covering the quotes and transactions from the major US exchanges (TAQ) grows exponentially, now at a rate of hundreds of terabytes per year. Although there are some parallel computing platforms that facilitate simple requests related to TAQ, there is nothing in place currently having the capacity to handle, more complicated things such as mergers between the TAQ and other potentially large databases.
The goal of this research project is to develop a deductive analytics platform to detect criminal activities leading to market destabilization by establishing the underpinnings of association or co-referencing for complex attributed or labeled graph structures that are derived from large heterogeneous data sources such as trading quotes, litigation releases, and public news.
To effectively cope with the extremely large financial data volumes, the dimensionality challenges, and the lack of effective research tools for large data we leverage the unique combination of data-intensive architectures from our recently granted NSF “MRI-R2: Acquisition of a Data Intensive Supercomputer (DISC) for efficient Deductive Analytics and the multidisciplinary team to innovatively solve problems in the area of financial forensics. Our research methodology draws upon methods from mathematical optimization, modeling, theorem proving, algorithm design, and distributed system implementation for the data processing and association activities. We are currently working towards the accomplishment of the following five key tasks in our research: (i) Identification and Selection of relevant data sources; (ii) Extraction of propositional graphs from textual data; (iii) Incremental Correlation Association algorithms for large scale data;(iv) Graph based algorithms and Information theoretic techniques for Relationship Discovery; and (v) Scalable in-DB analytics solutions for Deductive/Inferential analysis.