Home Page
Yahoo Charts
| |
GPF Release Notes (v1.6)
These notes come directly from Beta testing. New users should
first take a look at the FAQ and/or the
hints page.
Part 1
- Main new feature:
Patterns previously saved on the "Stats" sheet can now be reloaded.
Please note the genetic signature is used to restore GPF patterns along with
all other current (or default) settings, i.e. pattern type, bars back, etc...
are supposed to be those used during optimization. In some cases, it may be a
problem, but it can also be a good way to better analyze patterns. For
instance, one can test whether a small change in a setting (like slippage)
affects the quality of a pattern at run time. In other cases, the calculations
may be meaningless, like for instance if one tries to load type 3 patterns (IF
x AND y AND z THEN) but select type 2 patterns at run time. Some combinations
may even crash the GPF (unlikely though).
Please note that the chromosome description used to restore patterns must not
be moved (row 39 and below)
User notes can be added below the last chromosome value.
The "Load Saved Patterns" has been placed on the "Patterns & Trading" tab. If
valid patterns are found, the "Analysis" button is enabled to then calculate
stats, generate signals, etc... It is however obviously necessary to also load
data in memory ("Data Source" tab) prior to go any further.
Other notes:
- Signal information has been improved a little (Column N in
"Equity Curve" sheet")
For instance, entry can now look like "Signal Entry: /\ Next Bar's Open"
meaning entry long at next open.
To avoid confusion, if a subsequent entry occurs before the current signal
stops, for instance, if the signal is valid 5 bars ahead, and another signal
occurs after 3 bars, the exit signal of the first signal won't show.
However, the signal profit and loss is not updated accordingly. This means
that signals are still calculated individually.
- The "Equity Curve" sheet is renamed to accommodate the name
of the source data sheet. This will make it easier to compare when one uses
several data sheet for testing purposes. We remind you that out-of-sample
testing can be done either by holding a number of data records (at the end of
data file) or by using a separate data sheet.
One must however be careful as the pseudo equity curves are generated
individually. Nothing stops you from selecting a pattern on original data,
then loading new data, selecting another pattern, and try and compare two
completely unrelated patterns.
This change also applies to the "Trade Distribution" sheet.
- index has been adapted to include last (and current)
bar. Previously calculations stopped 'NumBarsAhead' bars up from end of file.
Of course, profit potential, or next Open, or any future value is unknown at
that stage.
Part 2
- There was still a small glitch when adding new data
at the end of the sheet. Signals were not calculated up to the current bar. It
should be fixed now.
I wish to remind you not to add lines in between quotes in the data file.
Pattern calculation would be affected.
You can however add a blank line (insert row) anywhere in your data file. Only
the first part will be read by the GPF. You can then optimise for patterns,
remove the blank line, load the whole data sheet, and run "Analysis" on the
complete data set.
- A bug has been introduced recently (block of code removed
inadvertently for testing purposes). It prevented usage of the "My Own
Objective" option (instead of the built-in calculated objectives). A column of
calculated objectives must be made available to the GPF (Data Source tab).
I wish to remind that all tunings of the distribution of such "output" data
(in neural terms) must be done beforehand. The GPF will use that data as is.
For instance, one cannot apply stop loss, noise, or outlier filtering again on
that data.
- On request, a GPF exit confirmation box has been added.
- When using stop/limit entries, the GPF generates a signal
that may or may not be filled on the following bar. The exact same signal
function is used during optimization, and later pattern generation. What can
therefore happen is that a signal is generated on the last available bar
during optimization, and disappears if not filled when new data is added. It
is quite normal.
- When the "Equity Curve" sheet is generated, depending on
the number of bars ahead, the same number of lines at the end of the sheet
will show zeros (Potential Gain/Loss column), as such information is obviously
not known at that stage. This applies to Next Open too, if that option is
used.
Previously, optimization stopped just before those lines, but prevented the
use of the GPF on current data (owing to the program design). see item 1
above. Adding zeros during optimization should have no or very limited
statistical impact, and does not bias profit calculations.
|