Finding time series among the chaos: stochastics, deseasonalization, and texture-detection using neural nets

Abstract

Summary form only given, substantially as follows. Problems of portfolio management have included several fundamental time-series problems. Parts of these problems are involved with the inevitable noisiness of financial data, parts with interactions and mode-locking among measures, and parts with the basic probabilistic nature of predictive systems in a rich environment. Modern neural networks have been used, to limited effect, to resolve them. Innovative techniques should prove more helpful. Among the fundamental issues for comprehending time series data are: (1) adjusting models dynamically, as errors emerge and corrections are identified; (2) promoting model-wide adjustment; (3) avoiding the tendency of least-squares forecasts to decay with time; (4) locating the range of plausible outcomes; and (5) complex prediction/correction optimization strategies. Techniques pioneered in neural networks have addressed each of these issues. The most common algorithms employed have been backpropagation variants. Recent advances in backpropagation make possible substantial improvements in identifying seasonality, modality and structural stability. Advances in recurrent networks allow context-sensitive adjustment of sharing and "elastic fuzziness", and new forms of reinforcement learning which permit the detection of interaction among dimensions and dynamic adjustment to that interaction. Reconstruction of priors and "deconstruction" of observer effects are also consequences of elastic fuzzy networks and dual heuristic programming.