Time Series Analysis


Time Series Analysis (TSA) is popular for estimating the process that underlies some output, or for forecasting from some observed behaviour over time. Much of the present research activity in the area stems from an inability to have distributed activity in the models used for process estimation or forecasting. Each new algorithm that is proposed may alleviate a problem in one area, while introducing a problem in another. The existing TSA methods seem to reward lack of knowledge of the process, as there is almost no way to assist the methods to obtain a better estimation if there is prior knowledge of the process - its limits, etc. This being so, the practitioners of TSA are not motivated to understand what it is they are estimating, and accept a "Black Box" result.

An active structure, which needs no rigid boundary between model estimator and model, provides a way to distribute activity throughout the model estimation process, and hence minimise its granularity.   The active structure can immediately use any information that is known about the process, because its base elements are the elements of analysis, a far wider set than is found in the typical TSA algorithm.

Existing Approaches

The existing approaches suffer in some way from granularity limitations. An algorithm requires a firm footing on which to stand, and as soon as that is delimited, the areas of inapplicability are also defined.

Current approaches are briefly described. There are several properties that should be looked for when assessing the methods:

Is the method recursive? If the method decomposes a time series into components, is the same method available on the components? If a seasonal component is identified, can the seasonal component itself be seasonal on a longer (or shorter) calendar.

Think global, act local. Does the method detect components at any scale, or, is the granularity of the method coarse? Methods requiring stationarity can be guaranteed to have coarse granularity.

Extensibility. How well does the output integrate into a dynamic and extensible model framework? Can linkages be easily established between components in separate time series analyses, or does the method suffer from the "little room" approach - the problem has been removed from its surroundings and taken to a mathematically small room, where it has been dissected under controlled conditions in isolation from its context. The results may be interesting, but if they are not easily put back into a larger framework, they are usually a dead end.

Identifiability of the components. Some methods, such as using a Fourier transform and turning a time series into many sinusoids, can represent the time series data very well, and yet provide not the slightest identification with the underlying process. The process has attributes - energy, inertia, randomness - are these attributes somehow visible in the model estimation.

Box Jenkins ARMA
The classic (1970s) Time Series Analysis approach uses a Box Jenkins algorithm to transform the data from a stationary process into Auto Regression and Moving Average components, the so called ARMA. The method allows estimation of actual pulses or inclusion of forecast pulses that do not fit the ARMA methodology. A tournament is used to select the best fit from 271 different possible models. Iteration on the parameters is then used to find increasingly better solutions.

Why not use this method? It doesn't work for a very large class of curves - it even shouldn't be used on the airline flights example used to introduce it because the variance is not constant over time. The differencing frequently used to make the data more amenable to analysis (effectively, taking the first derivative) further hides the connection to the underlying process. The projection cannot be bounded or controlled, and is not readily useable in other systems, or even in the same system for a different time series.

The wavelets method is used in image recognition. The object to be recognised may appear at a wide range of scales, so some scale invariant method must be used. It is an attempt to embed local context sensitive processing in the recognition process. The method can be used on non-stationary time series because cyclic detail at any scale can be recognised.

The method distances the results even further from the context than does ARMA. An image of an aircraft at one scale is not usually found buried in an image of an aircraft at quite another scale, but in TSA there may well be a connection between the two. Wavelet methodology for time series analysis assumes independence of components at different scales, or more precisely, has no way of linking the effects, as this would invalidate the premise of the mathematical analysis. One should expect complex interactions among the elements of a time series, but this method will not reveal them, except to the trained eye looking at the results of the analysis, which defeats the object. The results are not immediately useable in other systems.

Artificial Neural Networks
This approach relies on building layers of nodes, each connected to a preceding layer through weights. By adding more and more layers, and incorporating memory delay lines so previous inputs can be stored, the output of the network can come to resemble the time series. The method offers recursive local error elimination, which is good.

The method is poor on the grounds of identifiability - there will usually be no obvious point in the neural network that coincides with an obvious, and perhaps easily measurable, point in the process, so verifiability is almost non-existent. Processes that involve threshold switching (bouncing off a limit, say) will be very difficult to handle. Ignoring identifiability, there is no way to couple analytic knowledge about the process into the ANN - it has its own mechanism for operation, antithetical to analytic operators. It probably shouldn't be included under Time Series Analysis because there is no analysis, instead the time series is mimicked by a set of weightings. There is no way to couple other knowledge into the Neural Network.

A New Approach - Active Structure

All of the above methods are attempting to solve a distributed problem by either using simple mathematics or resistors. A time series may contain:

A slowly varying ramp
Seasonal variations
Cyclic variations
Level changes
Limit behaviour

The components will usually have an influence on each other, making them difficult to tease out into separate independent components, but this is what the various algorithmic approaches demand.

The active structure approach to TSA involves creating new elements at a basic level (analytic operators and variables), and combining these to form the different components of the time series, effectively exploding the information in the time series into analytic structure. This guarantees the finest granularity possible, and enables the interweaving of the components - the amplitudes of the elements of the seasonal component may be correlated with the base ramp.

It should be obvious that a system for process estimation and forecasting with even a trifling amount of context can do far better than the most sophisticated system with none. The active structure can be given as much, or as little, context as is available - that is, information on hard bounds on any component, probability known external to the series, etc.

What Is an Active Structure

It is a collection of variables, operators and links through which information flows in any direction. Information is propagated and stored in the links and at the variables. A change in the information in a link will cause an operator, to which it is connected and to which the change is "flowing", to become active and process the changed information. The operator will then usually propagate changed information in one of its links, including the link on which the incoming change occurred. Information flowing through the links can be a single values or a range. The activity in the network is "micro-scheduled", so there is no limit on granularity imposed by an external algorithm. The representation is extensible, merely by adding more operators, and it is identifiable. A known limit on the seasonal amplitude, say, can be added, ensuring that  predictions take into account any available information. An active structure can have sophisticated operations embedded in it. A projection variable can have a probability distribution generated by the analysis, and probabilistic relations can exist between the variables. If a projected value becomes known and there is no time to rerun the analysis, the relations will constrain the distributions of the still unknown projected values.

A Simple Example

We have a time series with a seasonal component, a ramp, and possibly some pulses (which may be valid events or measurement errors). The structure of the network near the time series values becomes:timedecomposition.gif (7538 bytes)

The Ramp variables provide a slow change in value by minimising the differences between them. The Seasonal variables themselves are split into a repetitive component and a pulse component through another PLUS operator.  The Pulse variables are there to accommodate divergences which the Ramp and Season variables cannot absorb. The structure can be used recursively - once the ramp and season have taken away their components, and level shifts have been accommodated, any remaining correlation in the Pulse values can be operated on in the same way.

This permits finding features at multiple scales and recursively minimising errors.

Correlations can be sought between the levels of processing, so a ramp on one set of seasonal multipliers may be correlated with the base ramp or the peak amplitudes of another seasonal influence. The resulting model, with correlation connections across levels, is far more indicative of the underlying process than an estimation process which assumes independence of the components.

The knowledge network approach offers:

Recursive error elimination
Cross linking of levels
Other knowledge, bounds etc., can be used to control prediction
Easy combination of result with other analytic structure - we forecast the ticket sales, now let's forecast the new planes we need

probsheet.gif (23035 bytes)

The diagram illustrates use of a probability surface to guide the projection. The surface is made up of polygonal sheets. In this case, the gray sheet represents 0% probability and the green sheet represents 50% probability. There is zero probability outside the gray sheet, so all of the projection must stay within it. The effect is to guide and control the projection as it is being produced, rather than crop the result.

The probability sheets emphasise the rich graphical working environment provided by the knowledge network approach. The analyst can examine all of the components of the time series, make connections with other knowledge to be used in the time series analysis, or use the result of the time series analysis in some other analysis, or the whole operation can be run in batch, still with connection to other knowledge.


A profligate use of analytic elements - thousands of them. However, we guarantee that fewer network operators are used than the number of neurons used by a human in "eyeballing" and seeking to understand a time series curve. The many elements allow a finer granularity and greater control than is available with coarser methods.

A Slightly Biased Comparison Scorecard

Method Box Jenkins Neural Network (ANN) Wavelets Active Structure
Identifiability 4 0 2 10
Extensibility 0 0 3 10
Recursive Error 3 10 0 10
Cross Linking 0 3 0 10
Other Knowledge 3 0 0 10

Time Series Analysis Windows

Some Aspects

Finding the theoretical seasonality surface.
Integration with Other Systems
The output of the analysis is lists of component variables, seasonality, pulses, etc., and a set of projection variables carrying probability distributions, and possibly relations among them. This form of output is readily integrable in other eCognition™ systems used for marketing of financial or insurance products or for risk assessment. Systems which cannot accept rich numerical input can be fed single values or triplets of values which contain limited probability information.
The pulse lists put out by Time Series Analysis can be used by Tupai's Data Mining to automatically search for correlated events.