As mentioned I am back to the drawing board.
I've trimmed my flipper code to just the bare minimum and a market filter to test different situations. I using backtester to generate all possible trades and then performing simple statistical analysis of various factors around the time of entry to see there is statistical significant predictors of the price. Backtester produces 20,000 trades and you can add whatever metrics you want and then copy the information to excel or statistical analysis program to investigate.
This is different from optimisation in a few ways. The main difference is you can optimise a parameter (e.g. only buy if close price <10) but you do not know if a 2% improvement in profit is statistically significant or just explained by chance.
For example. This is a plot looks at buy price and profit
I think it is pretty obvious that 1-10 is a sweet spot. Simple statistical testing (e.g. t-test or anova can confirm this).
I looked at a few things
1)Buy price
2)Delaying buy by a few bars post signal and looking at whether price drops are rises in this period
3)Current price relative to 100-period high.
After applying this to the flipper
On SP1500 i get
CAGR - 12% with max DD-28.55% over 1997-2011 period. (Data used from this period for analysis)
I then tested from 2012-now
CAGR - 14% with max DD 12.85%
and if i used a market filter from 2012-now
CAGR-10.91% with max DD 9.29%
Now whats interesting is that if I apply this to all ords results are equivalent (slightly better) but removing the extra conditions I added result in less than 5% chance in profit and draw down and you still have a profitable system.
But if you remove this extra conditions for the SP1500 your profits crash to 0% and drawdown 40%.
I think this adds to what
@peter2 mentioned about the us market.
Just for fun I applied that scan (Oct18) over my list of US optionable stocks (3600) and there were 1571 results. At 2 sec/chart it would take me almost an hour to go through. With distractions 2hrs.
This is why setting up trading systems for the US markets are so much more difficult. The number of available stocks is huge. The first task is to reduce the trading universe to a much more manageable number of stocks.
Reducing trading universe is potentially not only important to save time but profit.
On the note of time. The reason I didn't run the analysis on russell 3000 is that it when you also include historical constituents the analysis takes much longer (e.g. 30 seconds vs 90sec). It adds up!
(But a final run of new flipper on russell 3000 from 2012-2019 has CAGR of 13% and maxDD of 19% without filter and 12%/19% with filter on.
I would love if people continue with suggestions and ideas and I will try to test these. These results are weekly and it will be interesting to look at things on a daily basis as well at some point
Cheers
JB