Book Review: "Neural Networks in Financial Engineering: Proceedings of the Fourth International Conference on Neural Networks in the Capital Markets (NNCM-95)", by A.-P. N. Refenes, Y. Abu-Mostafa, J. Moody, A. Weigend

The papers in this volume apply to a wide range of topics in Computational Finance such as time series forecasting, volatility estimation, trading, and corporate distress. These are poorly understood problems that have defied accurate specification by even the most seasoned experts and have been hard to model in closed form, but for which significant amounts of data are available. The availability of data, coupled with nonlinear modeling tools such as neural networks, is making it possible to construct intuitively explainable behaviors of markets and to test them rigorously. This is essentially what the papers in the volume try to accomplish, although more rigor and extensive testing are required. Even though tools like neural nets allow us to construct complex nonlinear models, building accurate predictive models is difficult because financial markets are driven by a large number of economic and psychological factors, rumors, etc., that are not feasible nor desirable to construct retrospectively. Indeed, when examined in retrospect, markets exhibit a significant amount of “noise” that is hard to model. Nevertheless, by constructing the right kinds of inputs from the raw data and feeding these to a neural network, we can attempt to extract the residual structure in the markets. Many of the papers in the volume claim various degrees of success at prediction in equities and futures markets. However, a lot more work needs to be done in order to evaluate the practicality and scope of neural nets to forecasting and trading. For example, will the result on one currency hold up for another? Why or why not? Secondly, most of the reported results ignore transaction costs and slippage without which it is hard to evaluate results.