I had just about written off this approach, when I had another idea. Minor changes, and much *lower value* filter used to create the jump-delta boolean vectors. This puts a lot more information into each vector. And this improved results significantly. Plus, I "tuned" the training activity. This is network number 13, which gets 68.41 % accuracy on the main training dataset, and 58.14 % accuracy on the test data set. The results for the test dataset are shown here. Used conjugate gradient for direction, with line-search (Rudi's). The Xerion Neural Net technology can be used to create actionable results.

Neural-Network-AI Experimental Results: Developed portable Xerion + TSM + Lynx(ssl-enabled) + GNUplot platform on Linux (Fedora/Redhat) laptop platform, (ACER with Intel Centrino). This Linux laptop (Gnome Desktop) also runs current Firefox (modern gtk+2, glib, gdk, etc.). Wine - Windows emulator on Linux - is used to support a runtime-version of TSM, the Time Series data manager, which transforms raw price data into training cases for the Xerion-configured neural network (NN). For the current NN-driven AI under test, the training is sourced with boolean impulse-data from various daily market prices for tradable securities and commodities, for an 18 year period. The resulting neural-network can be evaluated for current datasets (ie. the last couple of weeks) on either this platform, or using an iPad or Android tablet.

[June 28, 2017] - Early Experimental Results: Screen shot showing Probability Calculator, Time Series Manager, (with linked GNUplot graphics), Xerion NN-AI (cmd-line mode runs GNUplot display, Xerion gui showing Hinton Diagrams of network unit values for most recent data case). The "plotValues" tcl/tk prgm shows boolean training target, and output of network boolean prediction in bottom, centre chart). All is integrated using the Fedora/RedHat platform, running on the dedicated AI box, an Intel 32-bit uniprocessor. Linux utilities "DOSemu" and "Wine - WINdows Emulator [or "Wine Is Not and Emulator"], used to support Probability Calculator app, WINE used to run Time Series Manager. Xerion was compiled from UTS source, with various minor modifications to support modern (sort of) Linux kernel (Fedora/RedHat Kernel #1 SMP. (Kernel is "old" now, but has a few custom bits compiled in)). Everything together at last, and running well. Results looking good - both technology, and market tone. Note that I modified the GNUplot display of "Actual" vs "Network Forecast" to show the predicted boolean output on the top (green line), with the actual training target on lower line. This makes it easier to see most-recent predicted network value, which can be expected to drive one's tactical market efforts. FD: I remain fully invested, long.