hscott
34 posts
msg #43978
- Ignore hscott |
5/21/2006 11:35:31 AM
heyen - On your predictor of what nasdaq will do 1 day hence I did more tests.
Here is what I did - took your filter
set{c1,count(day change 1 day ago above 0,1) + count(ind(qqqq,day change) above 0,1)}
set{c2,count(day change 1 day ago below 0,1) + count(ind(qqqq,day change) below 0,1)}
set{precaution,count(max(c1,c2) equal 2,100)}
precaution above 60
add column precaution
sort by column 5 descending
Using this filter I ran it with various date offsets up to 250 days back which is as far as SF would allow. I chose the top 4 picks from each test (with highest precaution) and inserted into the following filter I wrote
set{c4,count(ind(mgam,day change 1 day ago) above 0,1)+count(ind(pgnx,day change 1 day ago) above 0,1)}
set{c5,count(ind(shs,day change 1 day ago) above 0,1)+count(ind(npy,day change 1 day ago) above 0,1)}
set{c6,count(ind(qqqq,day change) above 0,1)}
set{c8,c4+c5}
set{c9,count(c8 equal 4,1) + count(c6 equal 1,1)}
set{c10,count(c9 equal 2,50)}
set{c11,count(c8 equal 4,50)}
add column c4
add column c5
and add column c6
and add column c7
add column c8
add column c10
add column c11
close above 200
This counts how many times all 4 select stocks went up as compared with how often the nasdaq rose the next day. Sometimes the 4 go up yet the Nasdaq falls so the predictor fails. c10 is the number of successful predictions while c11 is the number of times all 4 stocks rose. (close above 200 just reduces the number of matches to a smaller number). The test extends for 50 days. So to summarize, your predictor filter selects stocks based on a 100 day correlation with Nasdaq while my test uses that selection to go 50 days beyond your filters pick date.
Results are out of 24 times all 4 stocks rose the Nasdaq went up 14 times or 58%. This uses your filter applied to the whole universe of stocks. If you just apply your filter to Nasdaq however (as you did) the result improves to 23 of 33 or 70% as you were saying. There is an oddity. Of the 23 correct predictions 12 of them occurred during the test period from 50 to 100 days back. If you throw this period out you fall back to the 58% result.
Next I speculated that your filter give info that is based on a time span and we apply it to a future time span. Surely the more up to date the info is the better the results will be. So I shortened your filter to 50 days and my test filter to 25 days. So I use your picks for only 25 days after they are made. Results were indeed better - 22 of 29 for 76%.
So is it really true that we can now predict the sign of the change in Nasdaq the next day with 76% accuracy? (And maybe we could improve on that if we play around and fine tune it). Couldn't one make a lot of money just based on that knowledge? It doesn't really help me that much since I don't day trade but seems to me that you have indeed made a big discovery.
|
heyen
124 posts
msg #43998
- Ignore heyen |
5/21/2006 5:56:44 PM
hscott,
great work - 76% exceeds many trading systems!
Yes, there are surprising accuracies. The question i asked myself is:
how many days do we have to look at to avoid coincidence?
testing different time frames means a lot of manual work.
so thank you for your tests. i wish we could do a backtest.
but this seems to be beyond SF's actual capacities.
Good job hscott!
|
juha
52 posts
msg #44237
- Ignore juha |
5/26/2006 5:53:17 PM
hscott, could you explain your filter
sorry, I understood nothing from it :( Which variables c1-c11 has correlation with tomorrow's QQQQ?
thanks
Alex
|
shelupinin
120 posts
msg #44300
- Ignore shelupinin |
5/27/2006 7:17:23 PM
Hi !
I have another question: How to "backtest" such filters? All those filters "fetch" stocks which "predicts" tomorrow's QQQQ, so I need to buy QQQQ if most of those stocks with high "precaution" value goes up . How to set such condition in backtesting? By default "backtest" assumes to buy fetched stocks, not QQQQ if stokcs goes up !!! Please, somebody, help !!!
thanks in advance
|