Modelling, forecasting and trading with a new sliding window approach: the crack spread example. Issue 12 (1st December 2016)

• Record Type: Journal Article
• Authors: Karathanasopoulos, Andreas; Dunis, Christian; Khalil, Samer
• Is Part Of: Quantitative finance. Volume 16:Issue 12(2016)
• Page Start: 1875
• Page End: 1886
• Publication Date: 2016-12-01
• Subjects: Spread trading -- PSO RBF neural network -- MLP neural network -- HONN -- Sliding window training -- ARMA -- GARCH -- Threshold confirmation filters; Finance -- Periodicals; Business mathematics -- Periodicals; Finance -- Mathematical models -- Periodicals; Investments -- Mathematics -- Periodicals; Economics -- Periodicals; Finances -- Modèles mathématiques -- Périodiques; 332.015118
• DOI: 10.1080/14697688.2016.1211796
• Languages: English
• ISSNs: 1469-7688

Abstract

Abstract : The scope of this analysis is the modeling and the tracking of the crack spread with a sophisticated new non-linear approach. The selected trading period covers 2087 trading days starting on 09/05/2005 and ending on 21/12/2015. The proposed model is a combined particle swarm optimiser (PSO) and a radial basis function (RBF) neural network which is trained using sliding windows of 300 and 400 days. This is benchmarked against a multilayer perceptron (MLP) neural network and higher order neural network using the same data-set. Outputs from the neural networks provide forecasts for 5 days ahead trading simulations. To model the spread an expansive universe of 250 inputs across different asset classes is also used. Included in the input data-set are 20 Autoregressive Moving Average models and 10 Generalized Autoregressive Conditional Heteroscedasticity volatility models. Results reveal that the sliding window approach to modelling the crack spread is effective when using 300 and 400 days training periods. Sliding windows of less than 300 days were found to produce unsatisfactory trading performance and reduced statistical accuracy. The PSO RBF model which was trained over 300 is superior in both trading performance and statistical accuracy when compared to its peers. As each of the unfiltered models' volatility and maximum drawdown were unattractive, a threshold confirmation filter is employed. The threshold confirmation filter only trades when the forecasted returns are greater than an optimized threshold of forecasted returns. As a consequence, only forecasted returns of stronger conviction produce trading signals. This filter attempts to reduce maximum drawdowns and volatility by trading less frequently and only during times of greater predicted change. Ultimately, the confirmation filter improves risk return profiles for each model and transaction costs were also significantly reduced.