Case-based Algorithm for Pattern Recognition and Extrapolation (APRE Method)

The method developed and described in this paper departs from the traditional time series analysis approach. The starting premise is that a time series can be broken down into a number of characteristic cases, each of which potentially holds the key for indicating the value of the subsequent observation. The case that constitutes the beginning of the forecasting horizon (the reference case) is compared with all the past cases and the best-case match is identified. The differences between the best historical case and the reference case are used for predicting the first value in the forecasting horizon. The method is compared with some of the most prominent forecasting methods (exponential smoothing and Box-Jenkins ARIMA modelling). To validate behavioural properties of the original series versus the extrapolated values (forecasted series), measurements using the correlation (fractal) dimension are undertaken.