|
PureBytes Links
Trading Reference Links
|
Omega Man said -
> Dave,
>
> I hate to disagree with you because I like your
> name.
Sorry, but I'm not that crazy about your name Omega Man...
>But if you're wrong, you're wrong. And
> you're wrong. :-)
> This ATR exit is another LeBeau thing that makes
> no sense. First, you say: "Ok, I won't use a
> fixed stop or target since markets change over
> time..." Then, you proceed to set a fixed number
> of ATR's for your stop/target! Who says that
> swings in the S&P (or any other market) are still
> the same number of ATR's as they used to be?
> (They aren't - check into it.)
Now don't get excited, just pay attention :-)
In the first stage of system development I'm only interested
in answering my question, "What is the probability of this
entry signal making money" or, more specifically, "Does this
entry signal have a positive expectation of profit, and if
so, what is that expectation given a fixed ratio of risk to
reward?". I like this question because the answer will give
me a some numbers that mean something. And I can directly
compare these numbers to the numbers generated by testing
other entry signals.
Once I have selected an entry signal for testing the
simplest way to answer my question is to: 1) set a dollar
value for a money management stop, 2) set another dollar
value for a profit target, and 3) run the system and note
the percentage of winning trades. If I were building a
daytrading system and my test data only spanned the last
year or two then fixed dollar values for the mm stop (risk)
and profit target would probably be acceptable. But if I'm
running the test on long term data and the particular market
I'm running the test on has changed significantly over time,
as has the SP for example, then I need to take this into
account to arrive at a satisfactory conclusion. This is
where the ATR calculated exits come in.
To demonstrate, I ran the following simple test.
Data: SP500 futures continuous back-adjusted
Period: 04/21/1982 - 04/21/1999
Entry System: Simple MA crossover 4/20, (long trades only to
avoid SAR trades)
Trial one: $5000 mm stop and $5000 profit target.
Average # bars/trade over entire period = 38
Average # bars/trade in first year = 67
Average # bars/trade in last year = 2.5
Trial two: ATR exit system I posted, MM stop at 2 * ATR,
profit target at 2 * ATR (where ATR = Average(TrueRange,
25))
Average # bars/trade over entire period = 9
Average # bars/trade in first year = 9
Average # bars/trade in last year = 7.7
I didn't bother calculating #bars/trade values for all years
and measuring variation, but somebody could do that if they
thought it was relevant.
So what's the point? Just that by using the ATR calculated
exits the trades are being executed with some consistency of
behavior. This is what I want to see. Also, I don’t' want
trades stretching out for months at a time obscuring many of
the setups that I'm interested in testing.
Is there a variance in volatility over the test period? Of
course. Does this invalidate the results? Why would it?
Exactly the same ATR is used to calculate the risk and
profit target FOR EACH TRADE. So with the ratio of risk to
reward held constant for each trade, and with the real risk
and reward values constant as a percent of ATR, I have set
up the conditions to answer my question.
> A fixed dollar stop/target is no different than
> the "exit after so many bars" exit. Neither has
> anything to do with the market. Neither has any
> concept of market movement behind it. Both are
> invalid means for evaluating systems and will
> show the same profitability percentage (as Tim M.
> pointed out) for almost any entry. Why is that?
> You're ignoring Tim M.'s initial finding...
Are you really saying that all entry methods will show the
same percent winners and it is only the exit that determines
profitability or lack of it? This is a new one on me. Can
you present any evidence to convince me of this?
As to why Tim M found a consistent variation in winning
percentages when testing a variety of entry signals across
multiple markets, there could be several explanations. It
could be as Tim mentioned he was unintentionally testing the
same idea expressed in different ways. Or, it could be that
using the 5, 10, 15, 20 day exit system so muddies the
waters that differences between signals become
insignificant. I don't know enough about it to comment
further.
Btw, neither Tim M. nor I was talking about evaluating
systems, we were talking about evaluating *signals*
(entries). You might want to keep that distinction in mind.
> If you want to see every time a certain setup's
> criteria are met, use a PaintBar. But don't
> think that changing your exit criteria to
> something artificial so that you get entries that
> you wouldn't have gotten makes your system
> testing more "realistic". It doesn't.
Well, I don't care where the setups occur so why would I
want to use a paintbar? But I do want to make sure I capture
as many of the setups as I can in the test. The more events
included in the test the greater the confidence in the
results.
I don't understand why you think exiting on some multiple of
average true range should be "artificial". Why wouldn't this
be just as valid as any other exit technique? And again, you
seem to misunderstand the purpose of the test. I am not
trying to make the test more "realistic" - just trying to
answer the question.
> All it does is create entries that would not have
> occurred had you been trading the system.
I have no idea whether or not the entries would have
occurred if the system was traded because I am not testing a
system at this stage (see above). But by testing entry
signals separately from a complete system I'm recognizing
that the signal events can occur at any time in the future,
without regard to or dependance on what might happen after
the signal occurs.
In fact, I humbly submit that I have no idea what the market
will do after an entry signal occurs, and neither does my
system. I only know I've proven (as best I can) the trade
has a positive expectation of profit. Nothing more.
To complete the system I discard the risk and profit target
exits and move on to adding rules to maximize the profits
and minimize the losses.
Btw, when you discard an unsatisfactory system built with
your "holistic" approach how do you know you aren't throwing
out valid entry techniques that happened to be paired with
substandard exits, or vice versa?
Dave
Tradeworks Software
http://www.mechtrading.com
>
> The Omega Man
>
> ____________________________
>
> Dave DeLuca wrote:
>
> I found the testing method used by LeBeau
> difficult to evaluate. What if the entry signals
> look great with a 10 day exit, but terrible with
> the others?
>
> What I like to do is use a fixed money management
> stop against a fixed profit target. For example,
> test a system using a $1000 mm stop and a $1000
> profit target and examine the percent winners.
> Consider 50% to be no better than random. If you
> are building a trend following system then adjust
> the ratio, say a $500 mm stop vs. $1500 profit
> target. Now the random mark will be 25% winners.
> If it is significantly higher you might have something.
>
> The size of the fixed dollar stops and targets
> can be selected for each market depending on the
> market's contract value and normal trading range.
> But if the market's volatility changes
> significantly over the testing period (as in
> long-term back-testing of the SP)
> fixed dollar stops and targets are going to be
> misleading. In these cases I find it is better to
> use a method for measuring the average trading
> range in points and set stops and profit targets
> as a percentage of the average true range
> measured over the last 25
> bars. For example -
>
> Vars: Atr(0), RiskTgt(0), ProfitTgt(0), MP(0),
> RiskFactor(0.5), ProfitFactor(1.0); {2 to 1
> profit to risk ratio set here}
>
> Atr = Average(TrueRange, 25);
> RiskTgt = RiskFactor * Atr;
> ProfitTgt = ProfitFactor * Atr;
> MP = MarketPosition;
>
> if MP = 1 then
> begin
> ExitLong at EntryPrice + ProfitTgt limit;
> ExitLong at EntryPrice - RiskTgt stop;
> end
> else if MP = -1 then
> begin
> ExitShort at EntryPrice - ProfitTgt limit;
> ExitShort at EntryPrice + RiskTgt stop;
> end;
>
> What I am looking for is a simple validation of
> the entry signal. What is the probability that
> this setup will generate more profits than
> losses? The results are much easier to evaluate -
> "there is a 42% probability that this signal will
> make twice as much as it
> loses..." - not bad. Ignore the actual profits
> and losses in dollars.
>
> This method has some advantages over testing with
> programmed exit signals, especially in trend
> following systems. For one, it exposes a lot of
> entry setups that are hidden in those big trades
> where the system is letting the profits run. I
> want to hit every
> setup in the data and see how it turns out.
> Anything else is over-fitting imo.
>
> Dave
> Tradeworks Software
> Get a free random trade generator at
> http://www.mechtrading.com/tradestation/random.html.
>
>
|