The purpose of this paper is twofold. Firstly, to investigate the merit of estimating probability density functions rather than level or classification estimations on a one-day-ahead forecasting the task of the silver time series.
This is done by benchmarking the Gaussian mixture neural network model (as a probability distribution predictor) against two other neural network designs representing a level estimator (the Mulit-layer perceptron network [MLP]) and a classification model (Softmax cross entropy network model [SCE]). In addition, we also benchmark the results against standard forecasting models, namely a naive model, an autoregressive moving average model (ARMA) and a logistic regression model (LOGIT).
The second purpose of this paper is to examine the possibilities of improving the trading performance of those models by applying confirmation filters and leverage.
As it turns out, the three neural network models perform equally well generating a recognisable gain while ARMA benchmark model, on the other hand, seems to have picked up the right rhythm of mean reversion in the silver time series, leading to very good results. Only when using more sophisticated trading strategies and leverage, the neural network models show an ability to identify successfully trades with a high Sharpe ratio and outperform the ARMA model.