Time series forecasting is a challenge in many fields. In finance, one forecasts stock exchange courses or stock market indices; data processing specialists forecast the flow of information on their networks; producers of electricity forecast the load of the following day. The common point to their problems is the following: how can one analyze and use the past to predict the future? Many techniques exist including linear methods such as ARX or ARMA, and nonlinear ones such as the ones used in the area of machine learning.
In general, these methods try to build a model of the process that is to be predicted. The model is then used on the last values of the series to predict future ones. The common difficulty to all methods is the determination of sufficient and necessary information for a good prediction. If the information is insufficient, the forecasting will be poor. On the contrary, if information is useless or redundant, modeling will be difficult or even skewed.
In parallel with this determination, a suitable prediction model has to be selected. In order to compare different prediction methods several competitions have been organized, for example, the Santa Fe Competition, the CATS Benchmark Competition and the ESTSP’07 Competition.
After the competitions, their results have been published and the time series have become widely used benchmarks. The goal of these competitions is the prediction of the subsequent values of a given time series (3–100 values to predict). Unfortunately, the long-term prediction of time series is a very difficult task.
Furthermore, after the publication of results, the real values that had to be predicted are also published. Thereafter, it becomes more difficult to trust in new results that are published: knowing the results of a challenge may lead, even unconsciously, to bias the selection of model; some speak about ‘‘data snooping’’. It becomes therefore more difficult to assess newly developed methods, and new competitions have to be organized.
This special issue is based on extended version of papers presented at the joined ESTSP’08 (European Symposium on Time Series Prediction) and AKRR’08 (Adaptive Knowledge Representation and Reasoning) conferences. This shared event took place in Porvoo, Finland, from 17th to 19th of September, 2008. The goal of joining these conferences was to create an interdisciplinary forum for researchers who may widen their scope of attention beyond the usual scope of research. The cross-fertilization took place, for instance, by offering the attendees shared keynote talks. Prof. Marie Cottrell (Paris University 1) gave a talk on data analysis using Self-Organizing Maps. Prof. Jose Prı́ncipe (University of Florida) described information theoretic learning and kernel methods.
Dr. Harri Valpola (Helsinki University of Technology) explained how to extract abstract concepts from raw data using statistical machine learning methods. One specific shared theme of interest was anticipation, i.e., how an agent makes decisions based on predictions, expectations, or beliefs about the future. Anticipation is an important concept when complex natural cognitive systems are considered.