Investigators
Chenoweth Tim
Drossu Radu
Obradovic Zoran
Tomsovic Kevin
Vucetic Slobodan
Problem
Designing forecasting models for nonstationary time series whose characteristic parameters change over time is a challenging problem since often such time series can not be converted to semi-stationary series and so fitting a single model to such data is inadequate, while it might not be clear how many models is needed, which historical data is associated with a specific model and how to integrate multiple models for on-line foresting.
The problem of modeling financial markets behaviour is particularly difficult when modeling financial markets behaviour since such data is typically nonstationary and multivariate while still a lots of relevant information is unavailable to public. Developing appopriate foreacsting models for such domains in real-time is receiving much attention from researchers in industry and academia, but no general solutions are known.
Results
- Developing a library of historic models
We have recently developed an effective constructive learning algorithm for regime discovery in nonstationary time series (vucetic00). The approach is based on competition among regression models aided by averaging of their squared residuals over neighboring data points and incremental introduction of additional models when needed. - Identifying regimes in real-time
Starting from a partition obtained by the proposed algorithm a library of historically successful models can be established and used for on-line non-stationary time series prediction through an accuracy-based regime signaling mechanism developed in our previous work (drossu96bj), (drossu97a).
Here, three different forecasting scenarios are considered for deciding whether to reuse historically successful neural network models or retrain new ones when a change in the distribution is signaled (drossubook). - Embedding prior knowledge
For efficient on-line forecasting using non-linear models like neural networks we have shown that stochastic analysis can provide some initial knowledge about an appropriate NN architecture, parameter values, and data sampling rate for their rapid design (drossu96book), (drossu96aj),(drossu95c), (drossu95b). - Understanding electric power markets
The validity of the proposed techniques is evaluated in the context of understanding price vs. load relationships at California deregulated electricity markets where we show that a number of characteristic repetitative behaviours exist in the price time series and provide an analysis of each identified regime (vucetic01j). - S\"P 500 Index analysis
We have also proposed a multi-component system for analyzing more general interdependencies in financial markets. It shows promising forecasting results for the S\"P 500 and other very noisy and non-stationary phenomena in financial markets. In the preprocessing stage the proposed system determines relevant features for stock market prediction, removes some data as noise, and separates the remaining patterns into two disjoint sets (tim95j), (tim94). Next, two biased neural networks predict the market's rate of return, with one network trained to recognize positive and the other negative movements (timinpress), (tim96a). In the final stage, if appropriate, additional rule-based knowledge is added to the system (tim96bj), (tim95b), and a trading buy/sell recommendation is made by integrating all local estimates (tim96aj), (tim95a).
An extensive evaluation of daily and monthly trading results shows that the dual neural network system yields much higher return with fewer trades as compared to a more traditional single neural network predictor. In addition, the proposed system managed to achieve an almost twice larger annual rate of return when compared to that of the well known buy and hold investing strategy over a seventy month period.