|
PureBytes Links
Trading Reference Links
|
Hi Al, sorry to take so long to get back on
this. Thanks for the thoughtful test of the synthetic data and the
complete analysis provided. It is an interesting set of stocks you
selected, I presume the BOBE is either Bob Evans Restaurants or the symbol is
BBY for Best Buy - shouldn't make a difference either way, although I hope you
have been in BOBE for the past year or so.
I am not intimately familiar with ADX, however it
appears that your ATR constraint may have some impact on the signals that you
accept. At any rate that would likely increase the similarity of the
synthetic results with the actual instruments.
More importantly, the average directional movement
indicators, in my understanding, are oscillators that hope to exploit an
observation that prices trend upward by closing strongly upward; i.e. the
sentiment of the trend is still strong and buyers are interested in the
stock. Your random/scrambled buyer has no concept of interest (or
trend for that matter) and this property should not exist in the data. I
suspect the inconsistent metrics of the synthetic results indicate this missing
information. btw: I am using "trend" in a sense that requires some
historical feedback; I would not identify a dozen sequential heads asa
coin trend.
Reviewing your spreadsheet results, I am at a loss
for what I could conclude about the systems, other than there is little
correlation in system metrics. The average return of the synthetics is nearzero
with a large variation. The average win/loss are somewhat similar with roughly a
10% stddev. max and avg trade drawdown are consistent, however with a ATR stop
loss that would be expected.
You can certainly generate a large number of
synthetic baskets, however, I still fail to see the value in those. If you
somehow developed a profitable approach that worked for all random series, I may
expect it to work for non-random series - maybe - unless the system made use of
the metrics of the random series distributions, etc. You can indeed create
synthetics and test systems ad nauseam - but I see little to no value in that
effort.
You clearly did a lot of work and I do find the
results interesting. They do, however, support my suspicion thatthe
results of scrambled data are not terribly useful. Of course, I may have seen
what I expected! As we used to say with seismic maps, "if I didn't believe
it I wouldn't see it." Which brings another anecdote experience to
mind: the research center created a random seismic section, hung it on the
wall and listened to interpreters interpret the geologic significance.....
What do you conclude from the results? I have
no doubt that you can "see" trends and patterns in the random series, we have
all noticed similar "pictures" in by watching clouds roll by; however, I have
been unable to extract the value from such cloud formations.
Cordially,
Richard
<BLOCKQUOTE
>
----- Original Message -----
<DIV
>From:
Avcinci
To: <A title=amibroker@xxxxxxxxxx
href="">amibroker@xxxxxxxxxxxxxxx
Sent: Saturday, June 22, 2002 6:20
PM
Subject: [amibroker] Test of
Scrambler
Well, folks, I conducted a test of the Scrambler today. Here is a
description of the experiment. First, I chose 5 tickers that were not
correlated as to industry and put them in a watchlist. They were AAPL (Apple
Computer), AET (Aetna Insurance), ANF (Abercrombie and Fitch), BP (British
Petroleum), and BOBE (Best Buy). I then coded a trend-following system that
trades both long and short (ADX uptick for entry with a max 2.5-ATR stoploss
and a 4-ATR trailing stop, using Stephane's RemBuyTrail and RemShortTrail
calls as part of his RemBuy.dll plugin). I tested the system on those 5
tickers for the last 1000 bars (about 4 years, starting on July 1, 1998).
Then, I ran the scrambler to generate 1000 bars of synthetic data for each
ticker, creating a separate watchlist for those synthetic tickers. I tested
the system on those 1000 bars and recorded the results. I then replicatedthis
procedure 4 more times to generate 5 separate replicates of these randomized
synthetic tickers and therefore 5 independent forward tests of the system. The
results are summarized in an Excel spreadsheet attached to this message.
The column headings are self explanatory. The first row of data contains
the BACKtest of the 5 native tickers using real data from the last 1000 bars.
Keep in mind I used position sizing of 1% of capital (1R = $1000), so the %
return figures are necessarily small because of the way AB calculates % return
(on the basis of total equity). The system generated a small positive
expectancy of 0.076 over the last 1000 bars and a total net profit of $14,530.
The next 5 rows of data summarize the FORWARDtest results from the synthetic,
scrambled data. Data in the 2nd and 3rd rows were extremely close tothe
real data, giving about the same drawdowns, expectancy, net profit, avg. wins,
avg. losses, no. of wins and losses, etc. However, the next 3 rows showed
losses with negative expectancies and negative % returns. The drawdown numbers
were very close to the previous data. The biggest differences were in the%
profitable trades, i.e., the number of wins relative to the total no. of
trades. This was the main cause of the negative expectancies and returns.
Everything else seemed to be fairly uniform relative to the actual data.
When I viewed the synthetic data, I noticed that often the data were
highly volatile, much more so than usual. Since I was expecting this behavior,
I tried to control this somewhat by including in the buy and short statements
to trade only if the 20-day ATR was less than 7% of the closing price. So, all
trades generated were constrained by this filter. Perhaps I should have
imposed a 5- or 10-period moving average of the ATR being less than 7% ofthe
close to control volatility even more. Other observations included trends,
some lasting as long as several months, head-and-shoulder patterns, support
and resistance behavior, and consolidation periods. So, based on this
experiment, I see no reason why the scrambler cannot be used as out-of-sample
data to test a trading system. Now, having said that, I agree with the critics
who state that there is absolutely no past history to govern future price
behavior (i.e., trading psychology, supply and demand, etc.), and for that
reason use of scrambled data to test trading systems is somewhat unrealistic.
There's no doubt about that. However, I still contend it can be used as atool
for testing your trading system. I welcome any comments.
Al VenosaYour use of Yahoo! Groups is subject to the <A
href="">Yahoo! Terms of Service.
|