Why is it that a single initial back test is never as good as the subsequent Monte Carlo run.
Rename
What do you mean by "the subsequent Monte Carlo run"?
Here's an image of a dipbuyer backtest with $392K profit compared to 500 MC runs. The baseline backtest was about in the middle of the best and worst MC run, and, it was better than about 1/3 of them.
You can see the worst MC run was $325K profit. Never say never.
Here's an image of a dipbuyer backtest with $392K profit compared to 500 MC runs. The baseline backtest was about in the middle of the best and worst MC run, and, it was better than about 1/3 of them.
You can see the worst MC run was $325K profit. Never say never.
I run the "Run Backtest" first then Monte Carlo with "Trade Scramble.
The worst level in the Monte Carlo was about 500, the backtest was around 200. The Monte Carlo never got that low?
Notice that I used Same Date scramble. It's the only one I tend to use. Why? Because it gives you a distribution of probabilities for all the trades that your strategy would have taken on days where portfolio buying power was limited.
Other Trade Randomization methods tend to smooth out the returns. You picked Date Clustering. Unlike the Same Date scramble, clustering randomizes trade dates, but just to other dates that were traded.
For example, imagine your backtest used 10% equity and traded on 2 days. On July 5th it had 30 trade candidates and only July 19th it had 3. The baseline backtest only can take 10 [entry] trades, so you'll get 10 trades on July 5th and 3 trades on July 19th, assuming 3 of the first trades had closed.
But with Date Clustering, you're likely to get 10 (randomly picked) of all 33 raw trades on July 5th, AND, you're likely to get 10 more on July 19th. Now you're potentially looking at a completely new backtest with new raw trades that possibly didn't even exist in the baseline run.
Do that several thousand times, and maybe you can understand why the baseline run isn't even in the same range.
Other Trade Randomization methods tend to smooth out the returns. You picked Date Clustering. Unlike the Same Date scramble, clustering randomizes trade dates, but just to other dates that were traded.
For example, imagine your backtest used 10% equity and traded on 2 days. On July 5th it had 30 trade candidates and only July 19th it had 3. The baseline backtest only can take 10 [entry] trades, so you'll get 10 trades on July 5th and 3 trades on July 19th, assuming 3 of the first trades had closed.
But with Date Clustering, you're likely to get 10 (randomly picked) of all 33 raw trades on July 5th, AND, you're likely to get 10 more on July 19th. Now you're potentially looking at a completely new backtest with new raw trades that possibly didn't even exist in the baseline run.
Do that several thousand times, and maybe you can understand why the baseline run isn't even in the same range.
Thanks that resolves the matter.
Your Response
Post
Edit Post
Login is required