Back Testing results for Corn daily data
The corn data is very interesting for me. This price has definitely a cyclic nature. Look at the spectrum for corn prices calculated for 57 years of price history (1949-2006):
The
high narrow peaks indicate the presence of some cyclic processes in these data.
The strongest cycles are 3 months, one year and 2.5 years. It looks like the
annual and 3 month cycles are fundamental ones.
Applying methods of Chaos Theory, we can reveal the stochastic cycle which
represents a kind of
market memory in respect to previous price movements. The R/S analysis gives us
this:
The maximum on the yellow diagram shows that the period of this stochastic cycle is 2-5 years, i.e., this market "remembers" its history for 2-5 years. This is a very valuable information, it gives us a clue regarding the length of the training interval, how much price history we need to create the forecasting model. Besides, the maximum on the yellow diagram is localized, which means that the stock market memory is finite. All this information gives us a good hope regarding the possibility to create a forecasting model.
You will find below the results of express Back Testing provided for daily corn data. The applied models have been already back tested for Dow Jones Industrial index and Euro/USD. It helped us to speed up the process of Back Testing procedure, however this approach does not guarantee that the proposed Solutions are the best ones.
There are two suggested models: the red line represent the spectrum model and the blue line is for the dynamic model. Here you can see the outlook for these solutions calculated on 10 random intervals http://www.timingsolution.com/TS/BT/Corn/outlook.htm
The forecast horizon for this model is 10-25 price bars ahead (1 month). Recommended index is Relative Price Oscillator (1,25,25):
| MA1=Moving Average Period | 1 |
| MA2=Moving Average Period | 25 |
| MA3=Moving Average Period | 25 |
This is the Back Testing report:
| Model | dynamic_model.hpp | dynamic_model.hpp | dynamic_model.hpp | dynamic_model.hpp |
| NN Topology | 32 hidden | 32 hidden | 32 hidden | 32 hidden |
| Training Mode | 2000 before LBC train 15000 steps |
1000 before LBC train 15000 steps |
750
before LBC train 15000 steps |
500 before LBC train 15000 steps |
| +/- Statistics | +118 / -82 ChSq=3.2 | +131 / -69 ChSq=9.6 | +138 / -62 ChSq=14.4 | +130 / -70 ChSq=9.0 |
| Average (r,dev) | r=0.076 dev=0.0947 | r=0.117 dev=0.0935 | r=0.185 dev=0.0834 | r=0.164 dev=0.0542 |
| Model | spectral
nn model (sm=12.00;ov=15;min=7) |
spectral nn model (sm=12.00;ov=5;min=11) |
spectral nn model (sm=9.00;ov=32;min=9) |
spectral nn model (sm=9.00;ov=7;min=9) |
| NN Topology | 32 hidden | 32 hidden | 32 hidden | 32 hidden |
| Training Mode | 1000
before LBC train 15000 steps |
1000 before LBC train 15000 steps |
500 before LBC train 15000 steps |
500 before LBC train 15000 steps |
| +/- Statistics | +124 / -76 ChSq=5.8 | +110 / -90 ChSq=1.0 | +123 / -77 ChSq=5.3 | +117 / -83 ChSq=2.9 |
| Average (r,dev) | r=0.165 dev=0.0763 | r=0.044 dev=0.0780 | r=0.134 dev=0.0591 | r=0.134 dev=0.0699 |