The goal of this Cooperative Research project is to develop a novel system for time series analysis in a user-friendly environment, featuring newly developed algorithms and hardware. By user-friendly environment we will mean the intelligent interface, based on multicriteria analysis approach, allowing a non-mathematically educated user to model a series of discrete events as a time series, investigate its properties, forecast future events, and provide assistance by a menu guided help system.

The system will consist of a Kernel, which will be universal tool for a large class of applications, specialised modules, serving specific applications, and a user-friendly interface. The kernel will contain classic time series analysis algorithms, such as Kalman filter and prediction as well as novel methods, developed within the proposed research project, such as :

·         Monotonic aggregation methods, which were developed by the SME Co-ordinator jointly with the Department of Operations Research, University of Mining and Metallurgy. These methods of time series analysis are particularly useful for financial time series analysis, like Forex quotations, as well as for medical time series, such as encephalography. The results are independent from any averaging on the preceding time interval and reflect the real na­tu­re of the process.  In turn, the mean number of aggregated runs of lower order per run, considered as a function of time, may serve to determine the depth of the Moving Average process.  Monotonic aggregation returns function, which contains salient information about series. These data can be processed in further stages in different ways – by neural networks or statistical decision theory methods.

·         It has been discovered that if the Kalman filtering can be coupled with runs analysis then the computational efficiency of the filtering drastically increases.  This will make possible applications of filtering to simultaneous tracking of multiple vehicles, analysis of complex vector time series occurring in medical diagnostics etc. Moreover, pattern analysis methods will be useful for recognising of characteristic formations in the state space, improving thus the performance of short-time forecasting. The latter methods were developed recently at the University of Mining and Metallurgy, which is supposed to be a subcontractor in the Step 2 project.

The software will be developed as a set toolboxes for specific applications.  Below we describe two applications, which appear to be most characteristic and having a market potential :

 

Middle- and long-range weather forecasting

The weather characteristics can evidently be considered as a vector time series. The weather state at each moment of time can be modelled as a function: ³®Â^N , which assigns N-dimensional vector of real parameters (like temperature, pressure, humidity, their gradients) and symbols from a K-element vocabulary (cloud types) to every point of 3-dimensional space.  The first problem related to the above touches upon the data acquisition. Some of weather parameters can be retrieved from ground-level measurements, some other can be obtained form satellite images. The images can be interpreted by specialists and the resulting description input into our system. However, we plan to develop a module for automatic interpretation of satellite images. The most difficult part of it would be cloud classification. Each cloud type would be implemented as a model object, i.e. the set of procedures integrated with data, which allows identifying related object in the real image. For successful and robust identification every model has to join different approaches and use all the information which can be recover from a satellite image. We plan to apply texture analysis, skeleton classification and structural approach – decomposition to lower-level objects, which can be classified using statistical decision theory methods. As far as we know neither automatic cloud recognition nor weather forecasting using symbolic dynamics has been implemented yet in other systems.

 

    Decision support system for the Forex trade

This module will be developed to support decision making in Forex trading.  It is designed to generate sell/buy/hold signals and the allocation size in each traded currency separately.  The first task is carried out by means of carefully selected forecasting models.  The novelty of our approach relies on a suitable time series pre-processing and in optimisation of these models in order to achieve more accurate prediction.  The second task concerns one of the most important problem in finance: consumption/investment problem  We propose a new method of determining the size of allocation based on some measures of predictability (or randomness).  The system consists of three submodules.  Each of them has its own set of parameters, which will be optimised according to predefined goal function.  The forecasting submodule will apply two families of forecasting models.  The first group includes most frequently used by FX traders technical indicators. Models in second family apply neural networks and pattern recognition theory.  There is a significant difference between these groups.  Technical indicators usually give us a direct sell/buy signal while the methods from the second family generate a price forecast, which has to be transformed to sell/buy signal.  This transformation is implemented in the strategy submodule.  An input to this module is an output from the forecasting module, i.e. a preliminary decision d_M (sell/buy/hold), which can be modified due to some predefined rules.  These rules, which we call strategies are fixed.  The basic strategies used by real-life traders are “stop loss” and “take profit”.  The system will allow to implement pre-defined or user-defined more complicated strategies. The remaining Allocation submodule will generate an additional decision variable c(t) which can be interpreted as a size of capital allocation in time t.

 

Parameter optimization scheme

The other potential application include :

·       Vehicle tracking using Kalman-filter-based modelling approach

·       Microrobot path planning in a moving environment, using Kalman filtering and ARIMA-based methods as well.

The software will be developed as a C++ application with user-friendly interface.  For commercial purposes building stand-alone Windows application for Windows 2000 or Unix will be necessary. The user wilI be able to modify his/her preferences regarding the forecasting rules and use learning algorithms implemented within the system.  At each stage of the procedure it will be possible to export the results for further processing by other forecasting systems.

The structure of the anticipated partnership

Task name	Task acronym	AMSOFT Ltd.(co-ordinator)	SME2 (path planning)	SME3 (GPS/GSM programming)	Universities (subcontractors)
Co-ordination	Co-ordination	+
Kernel of the system (C++)	Kernel	+		+	+
Using aggregated functions to time series forecasting  – prototype version	Aggregation1	+			+
Monotonic aggregation procedures (C++) based on existing prototype software	Aggregation2	+		+
Module for time series forecasting using aggregated functions (C++)	Aggregation3			+	+
Pattern library – prototype version	Patterns	+	+		+
System for retrieving weather data (C++)	Weather		+	+	+
Software tests	Tests	+	+	+	+
Final report	Report	+	+	+	+

Allocation of tasks among enterprises and subcontractor is presented in Table 1 above.

The SME Coordinator, AMSOFT Consulting Ltd., Kraków, Poland, will itself develop a portion of the software system for financial time series analysis, the user interface of the decision-support system based on multicriteria decision-making approach, will be responsible for subcontracting and the data and research results exchange among the partners.  Since its software products are usually further developed and distributed by software vendors, it will offer its experience to all partners during the product commercialisation phase as well.

The Microrobotix s.a.s. is a spin-off of the University of Naples, will further develop and apply a robot tracking and path-planning system for their innovative small-size robots. This will increase their market share and extend the company’s expansion area. 

The ZUE Elsta, Wieliczka, will implement the Kalman filter based tracking module developed within the project to vehicle tracking in their GPS/GSM system and will give continuous feedback on the software performance.

We seek two to three further SME partners to be included into the CRAFT consortium.

Total Project duration : 24 months

Total project costs : 2 M €

Total EU funding: 1 M €

Funding sought : 1 M €

Research resources needed : algorithms for dynamic time series analysis, user-friendly interface design, satellite data

 

Economic impact and exploitation potential

Computer-aided time series analysis has a very wide range of possible applications. Those mentioned above seem to be particularly important, but the kernel module of the software system may serve as a base for the development of further specialised modules.  In spite of great importance of time series analysis in economy and finance, shown in the interest of banks, consulting companies and portfolio investors, existing tools are often insufficient for real-life sophisticated applications. We believe that the final product resulting from the project presented will help to overcome problems related to nonlinearities, non-stationarity, long-memory, persistence and other phenomena in real-life time series.  The development of the system here presented will increase the competitiveness of all the SME partners on the global market. An important advantage of planned system is a possibility of future using the time series analysis kernel for further application, which contributes to the multiplicative effect of the proposed research project.