Cumulate orders and decimal position size strange backtest results
-
OK. Let’s take this simple system that reproduces the whole mess. It runs in the DAX min 1 minute chart, 1 EUR mini contract, :
//Test system with integer position sizes (0 decimals), DAX 1 EUR Mini contract, 1 minute chart number = 1 + ((strategyprofit + 10000) / 125) n = round(number) If n <= 1 then n = 1 endif b = 0 If a = 0 and barindex = BIin + 20 then buy n contracts at market a = 1 b = 1 BIin = barindex endif If a = 1 and b = 0 and barindex = BIin + 20 then sellshort n contracts at market a = 0 BIin = barindex endif set target pprofit 10 set stop ploss 10This system alternately buys and sells short a position every 20 bars. No position therefore lasts longer than for 20 bars.
Now, let’s take the same system, but replace integer position sizes by position sizes with 2 decimals only by using n = round(number * 100) / 100 instead of n = round(number).
//Test system with decimal position sizes (2 decimals), DAX 1 EUR Mini contract, 1 minute chart number = 1 + ((strategyprofit + 10000) / 125) n = round(number * 100) / 100 If n <= 1 then n = 1 endif b = 0 If a = 0 and barindex = BIin + 20 then buy n contracts at market a = 1 b = 1 BIin = barindex endif If a = 1 and b = 0 and barindex = BIin + 20 then sellshort n contracts at market a = 0 BIin = barindex endif set target pprofit 10 set stop ploss 10It should yield exactly the same number of positions, lasting for 20 bars or less, but it DOES NOT.
The backtest curves superficially look quite identical, but in detail, they are not at all. Test done for 10.000 bars.
[attachment file=57396]
Now, let’s compare the backtest reports :
[attachment file=57397]
Both backtests should yield the same number of positions and the same time in the market. However, the system with decimal position sizes has more positions (625 instead of 526 for integer position sizes), and it seems to be much longer in the market, 92.51% instead of 69,87% for integer position sizes.
Now we will see, why this is so :
[attachment file=57398]
As in your system, Vonasi, the test system with decimal position sizes shows positions in the backtest report that do in fact not exist, and they last longer than they could in fact (longer than for 20 bars).
Now let’s take the longest wrong position with decimal position sizes (40 bars on 28.12.2017, 15:02 to 28.12.2017, 15:42) and see what the chart shows during this time :
[attachment file=57399]
Both systems show the same actual positions, the “wrong” position with 40 bars is not there, but instead we see once again these strange “spikes” around the time when the wrong position was reported for decimal position sizes.
So, I conclude that this is a general and reproducible bug for decimal position sizes, and I ask PRT to correct this, please.
I attach .itf files for both systems in the next post (this one is full).
This problem is still happening when forward testing in demo in any strategy using decimal position sizing and cumulative orders.
I noticed another strategy today with the same odd behaviour. It believes a position is open when in reality there is none. Image attached.
So that is two weeks of forward test results that are completely useless and a waste of time. I will have to start again with a fixed position size. Anyone testing in demo with decimal position size and cumulate orders should do the same as you may believe you have a fantastic result when in reality it is a load of rubbish.
- You must be logged in to reply to this topic.
Summary
This topic contains 38 replies,
has 5 voices, and was last updated by 
8 years ago.
Topic Details
| Forum: | ProOrder: Automated Strategies & Backtesting |
| Language: | English |
| Started: | 12/22/2017 |
| Status: | Active |
| Attachments: | 17 files |
