Emerging Missing Data Estimation Problems One drawback of concept drift is that for a high volume of non-stationary data streams, the time it takes to predict may grow indefinitely (Last, 2002). In all cases of concept drift, incremental learning methods that continuously revise and refine the approximation model need to be devised and these methods need to incorporate new data as they arrive. This can be achieved by continually using recent data while not forgetting the past data. However, in some cases, past data might be invalid and may need to be forgotten. Harries and Sammut (1988) have developed an off-line method for partitioning data streams into a set of time-dependent conceptual clusters. Their approach was, however, aimed at detecting concept drift in off-line systems. This work looks at a technique of detecting concept drift in an on-line application. CONCEPT DRIFT DETECTION USING HETEROSKEDASTICITY Techniques of detecting concept drift are quite essential in time series data. The biggest challenge to this task is due to data being collected over time. Ways of detecting concept drift may vary in accordance to the pattern at which the concept is drifting. In most cases, the use of a window, where old examples are forgotten has proven to be sufficient (Last, 2002; Helmbold & Long, 1991). Known examples of window based algorithms include Time-Window Forgetting, FLORA and FRANN (Widmer & Kubat, 1993). A cyclically drifting concept exhibits a tendency to return to previously visited states. However, there are many algorithms such as STAGGER (Allison, 2002) and FLORA 3 (Widmer & Kubat, 1993)