what the hell happened!?!?

Quote from feng456:

i'm looking at those excel files and i really don't understand what's going on. please help?
What is this? A CEO talking to a quant about his CDO pricing pre-2008?
 
my data was backtested manually. what i have is basically on list of 4 years of trades on excel done manually (yes it took a long time). it's a list consisting the following format:

June 2011
1st 5
2nd -5
3rd 5
.
.
.
Total for month: 10
etc.etc.


how do i do a monte carlo analysis on this? i have absolutely no idea because i never learned how to do it so thats why i was asking for a site where i could learn it. i looked it up on google and so far the tutorials ive gotten have no connection to my dataset.
 
Quote from feng456:
how do i do a monte carlo analysis on this? i have absolutely no idea because i never learned how to do it so thats why i was asking for a site where i could learn it. i looked it up on google and so far the tutorials ive gotten have no connection to my dataset.

There are many variations of MC techniques, but the basic idea is to generate synthetic equity curves by resampling the equity curve generated from running your system on historical data.

Here's one simple way you might do this:

1) Run your system on historical data and compute your daily returns. For a four year run this generates ~1000 samples.

2) Generate a synthetic 12mos equity curve by randomly chosing 252 samples from your historical returns in step #1. Compute the yield and max drawdown and save it.

3) Repeat #2 many times (>1,000). This generates a distribution of synthetic 12mo returns and drawdowns that you can use to calculate confidence intervals and the like.

Note that this method will scramble out any serial correlations that might exist, which will make your drawdowns look less severe than they might actually be. But you get the idea.

Here's a paper from the group I mentioned that describes the technique in more detail:

http://www.tradingblox.com/Files/MC_resampling_Nbars.pdf
 
Quote from jazzguysoca:

There are many variations of MC techniques, but the basic idea is to generate synthetic equity curves by resampling the equity curve generated from running your system on historical data.

Here's one simple way you might do this:

1) Run your system on historical data and compute your daily returns. For a four year run this generates ~1000 samples.

2) Generate a synthetic 12mos equity curve by randomly chosing 252 samples from your historical returns in step #1. Compute the yield and max drawdown and save it.

3) Repeat #2 many times (>1,000). This generates a distribution of synthetic 12mo returns and drawdowns that you can use to calculate confidence intervals and the like.

Note that this method will scramble out any serial correlations that might exist, which will make your drawdowns look less severe than they might actually be. But you get the idea.

Here's a paper from the group I mentioned that describes the technique in more detail:

http://www.tradingblox.com/Files/MC_resampling_Nbars.pdf

Excellent post. Really great stuff, I hope folks realize what your post contains. I've been working on serial correlation as well (random sampling, lag etc). Data that is serially correlated and is not randomly sampled will always yield optimistic results compared with results that have been randomly sampled with sufficient space between samples (lag) to achieve reasonable independence.

I ran into this when I was working on some radar tracking data for a client. Closed-loop tracking system errors are not independent or are any outputs from a Kalman filter.

Maybe the best post I've seen on ET, certainly in the Top-10. Thanks.
 
Sorry to hear. And somewhat frightening. But I'm sure it will never happen to me. :eek:

I'm curious about the frequency of trades. How many per month?



Happy Solstice Holidays everyone.
 
Quote from total_keops:

What is this? A CEO talking to a quant about his CDO pricing pre-2008?

Awesome, sir. Just plain awesome :D
 
Quote from 377OHMS:

Excellent post. Really great stuff, I hope folks realize what your post contains. I've been working on serial correlation as well (random sampling, lag etc). Data that is serially correlated and is not randomly sampled will always yield optimistic results compared with results that have been randomly sampled with sufficient space between samples (lag) to achieve reasonable independence.

I ran into this when I was working on some radar tracking data for a client. Closed-loop tracking system errors are not independent or are any outputs from a Kalman filter.

Maybe the best post I've seen on ET, certainly in the Top-10. Thanks.

Thanks for the kind words, ohms - appreciate it!

For those interested in more techniques along these lines, try googling "White's Reality Check" (there's a WRC paper on the reddit group I mentioned as well).
 
Quote from feng456:

I designed my leverage under the assumption that the worst drawdown in the backtested 4 years (prior to now) was the worst it could get so now since it's much worse, I am getting wiped out.

Well I guess you learned a lesson. The past is not indicative of the future. Chalk it up and move on...
 
Back
Top