Data analysis. To analyze the data, we computed: (1) The pairwise correlation coefficients of the 35 technical indicators. (2) The maximal pairwise correlation coefficients of all indicators and all linear combinations of two indicators. This analysis revealed that only 4 indicators are not highly correlated. For such reasons, our nets see only the 8 most recent DAX-changes and the following technical indicators: (a) the DAX value, (b) change of 24-week relative strength index (``RSI'') - the relation of increasing tendency to decreasing tendency, (c) ``5 week statistic'', (d) ``MACD'' (smoothened difference of exponentially weighted 6 week and 24 week DAX).
Input data. The final network input is obtained by scaling the values (a-d) and the 8 most recent DAX-changes in . The training set consists of 320 data points (July 1985 to August 1991). The targets are the actual DAX changes scaled in .
Comparison. The following methods are applied to the training set: (1) Conventional backprop (BP), (2) optimal brain surgeon / optimal brain damage (OBS/OBD), (3) weight decay (WD) according to Weigend et al., (4) flat minimum search (FMS). The resulting nets are evaluated on a test set consisting of 100 data points (August 1991 to July 1993).
Performance is measured like in section 5.3.
Results. Table 4 shows the results. Again, our method outperforms the other methods.
|
Parameters:
Learning rate: 0.01.
Architecture: (12-9-1).
Training time: 10,000,000 examples.
Method specific parameters:
OBS/OBD:
.
FMS:
;
.
If
then
is set to 0.03.
WD: like with FMS, but .
See section 5.6 for parameters common to all experiments.