Neural networks for financial forecasting
Neural networks demonstrate great potential for discovering non-linear relationships in time-series and extrapolating from them. Results of forecasting using financial data are particularly good [LapFar87, Schöne90, ChaMeh92]. In contrast, traditional statistical methods are restrictive as they try to express these non-linear relationships as linear models. This thesis investigates the use of the Backpropagation neural model for time-series forecasting. In general, neural forecasting research [Hinton87] can be approached in three ways: research into, the weight space, into the physical representation of inputs, and into the learning algorithms. A new method to enhance input representations to a neural network, referred to as model sNx, has been developed. It has been studied alongside a traditional method in model N. The two methods reduce the unprocessed network inputs to a value between 0 and 1. Unlike the method in model N, the variants of model sNx, sN1 and sN2, accentuate the contracted input value by different magnitudes. This different approach to data reduction exploits the characteristics of neural extrapolation to achieve better forecasts. The feasibility of the principle of model sNx has been shown in forecasting the direction of the FFSE-100 Index. The experimental strategy involved optimisation procedures using one data set and the application of the optimal network from each model to make forecasts on different data sets with similar and dissimilar patterns to the first. A Neural Forecasting System (NFS) has been developed as a vehicle for the research. The NFS offers historical and live simulations, and supports: a data alignment facility for standardising data files with non-uniform sampling times and volumes, and merging them into a spreadsheet; a parameter specification table for specifications of neural and system control parameter values; a pattern specification language for specification of input pattern formation using one or more time-series, and loading to a configured network; a snapshot facility for re-construction of a partially trained network to continue or extend a training session, or re-construction of a trained network to forecast for live tests; and a log facility for recording experimental results. Using the NFS, specific pattern features selected from major market trends have been investigated [Pring8O]: triple-top ('three peaks'), double-top ('two peaks'), narrow band ('modulating'), bull ('rising') and recovery ('U-turn'). Initially, the triple-top pattern was used in the N model to select between the logarithmic or linear data form for presenting raw input data. The selected linear method was then used in models sN1, sN2 and N for network optimisations. Experiments undertaken used networks of permutations of sizes of input nodes (I), hidden nodes (H), and tolerance value. Selections were made for: the best method, by value, direction, or value and direction, for measuring prediction accuracy; the best configuration function, H - I 4), with 4) equal to 0.9, 2 or 3; and the better of sN1 and sN2. The evaluation parameters were, among others, the prediction accuracy (%), the weighted return (%), the Relative Threshold Prediction Index (RTPI) indicator, the forecast error margins. The RTPI was developed to filter out networks forecasting above a minimum prediction accuracy with a credit in the weighted return (%). Two optimal networks, one representing model sNx and one N were selected and then tested on the double-top, narrow band, bull and recovery patterns. This thesis made the following research conthbutions. • A new method in model sNx capable of more consistent and accurate predictions. • The new RTPI neural forecasting indicator. • A method to forecast during the consolidation ('non-diversifying') trend which most traditional methods are not good at. • A set of improvements for more effective neural forecasting systems.