Sample Performances

We have backtested various models and plotted the Log Absolute Fractional Error (LAFE) on the y-axis below(obtained by normalising the logarithm of the absolute fractional error). The red curve  shows the prediction error of symmetry based FMI models. These have been tested against the best neural networks that the Princeton Computer Science Department could provide us (blue and green denoting random forest and Bayesian respectively) versus the best performing model out of the generalized ARIMA family (in yellow). The samples are randomly selected stocks from the S&P 500 with basic time series data supplied by Yahoo Finance. As expected, with this limited information, most methods have significant prediction errors. FMI has a lower error rate than others although even that may not be enough to trade depending on trading costs and slippages.  As the data is enriched 'trading signals' strengthen and stabilize, and can be used for increasingly profitable trading strategies.