1M backtest available… how do you manage it?
- This topic has 8 replies, 8 voices, and was last updated 3 years ago by
.
-
hmmm…. initially i just thougt, yeah, run it all. But maybe….
1. Code concept
2. Backtest on last know “current“ market conditions. In this case, post US election and tweetgate. Nov 3+
2.5 optimise
3. When good, walk forward 70/30
4. If still good, VRT test
5. if still good, run it on all bars available
5.5 rerun walk forward, VRT
6. if still good, forward test with demo (optional for me 😬)
7. If still good, order wife new handbag and me new golf clubs.
Something like that, at least on paper anyway!
1 user thanked author for this post.
Are you trying to create a strategy on the whole million? That greatly expands the backtesting required time and difficulty but allows a theoretically more solid strategy.
Or are you trying to use a smaller number of units? On which criterion?
I think that optimizing parameters on 1 million bars will firstly result in a further curve fit. Which does not guarantee future performance at all. Optimizing on 1 million bars means in most cases searching for the perfect algorithm that works under all market conditions : High and low volatility, market crash and a DAX price of 5.000 or 14.000. I don’t think that such an algorithm exists that will yield good results in out-of-sample conditions.
1 million bars can be useful for time scales below 1 minute. For example, in a 10 second chart, you can backtest over many months now instead of 7-10 days only. Any backtest and curve fit has a broader statistical basis, then. And you can test this curve fit immediately out of sample in earlier time periods. For example, I have reoptimizied several of my 10 second algos over 1 million bars from March, 2020 to January, 2021 (rather high volatility). Backtesting these curve fits from 2015 – February 2020 (entirely out of sample, conditions of high and low volatility) showed that almost no algo yielded positive results in these periods. Most of them produced losses in the order of number of positions * spread. So, 1 million bars applied correctly can help to sort out even more non-performing algorithms.
A few of my older algos retested over 1 million bars (mostly the whole time span from 2010-2021, no low time scale) have surprisingly shown quite good out of sample results. With these algo types I continue working right now.
This has been my experience with 1 million bars, so far. And 1 million bars make you think a lot about the efficiency of your codes and parameter optimizations, because when they are too complicated, they will take an eternity to run.
It only goes back to Aug 2010 as that’s all the tick by tick data they’ve got. All TFs of 5 min and over will max out at that point. Otherwise it’s proportional – 2min is about 6 years, 1min is ~3 years etc.
1 user thanked author for this post.
Find exclusive trading pro-tools on 
