Forecasting on complex datasets with association rules

Forecasting in complex fields such as financial markets or national economies is made difficult by the impact of numerous variables with unknown inter-dependencies. A forecasting approach is presented that produces forecasts on a variable based on past values for that variable and other, possibly inter-dependent variables. The approach is based on the intuition that the next value in a series depends on the last value and the last two values and the last three values and so on. Furthermore, the next value depends also on past values on other variables. No assumptions about the form of functions underpinning a dataset are made. Rather, evidence for each possible next value is collected by combining confidence values of numerous association rules. The approach has been evaluated by forecasting values in a hypothetical dataset and by forecasting the direction of the Australian stock market index with favorable results.